Chain of Thought | AI Agents, Infrastructure & Engineering

What breaks when agents become the customer? Most of the integration layer, APIs, SDKs and webhooks alike, was built for human developers and for servers that are always listening. Agents are neither.

When an agent runs on a laptop, closing the lid takes down the tunnel its webhooks come through. When the lid opens, it gets a storm of retries from 12 hours earlier. Svix founder and CEO Tom Hacohen's answer is to let the agent poll for what it missed.

Svix delivers webhooks for companies like Clerk and Resend, and it created Standard Webhooks, the spec OpenAI, Anthropic and Google Gemini have adopted. In this episode of Chain of Thought, Tom joins Conor Bronsdon to talk through designing APIs for agents that adapt, whether agents still need SDKs, and why he wants less code even when agents write it. He also introduces Poke, Svix's new spec for making event-driven APIs work for agents.

This episode is sponsored by Svix.

We cover:
  • Why events are harder for agents than API calls: the sender controls them, and they arrive outside the agent's session
  • How polling endpoints and Svix's plugins for Hermes and OpenClaw let an agent pick up where it left off
  • Why Tom calls JSON an abomination and webhooks duct tape, and why worse is better anyway
  • Why Svix created Standard Webhooks when everyone implemented webhooks differently, and how the spec let Kong and ngrok add webhook verification
  • Whether agents still need SDKs, and why Svix's skills work like a solutions engineer while its docs stay written for humans
  • Why Tom wants less code even when agents write it, and what piling patches on patches costs
  • Poke, Svix's early spec for letting agents subscribe to and trigger events across services, released so feedback can shape it while it's still malleable
Chapters:
(0:00) Duct tape, retry storms and the case for simplicity
(1:20) Why integrations got easier and harder at once
(3:09) How EteSync's webhook requests led to Svix
(5:28) Worse is better: why webhooks persist
(7:21) What changes when an agent receives the event
(9:16) What happens when the laptop closes
(11:16) Hermes, OpenClaw and agents that poll
(13:13) Letting agents choose their own subscriptions
(14:52) The long tail of the web without APIs
(19:57) Why Svix created Standard Webhooks
(24:34) How Standard Webhooks is governed
(26:33) Following real usage, including MCP
(28:34) Designing APIs for agents that adapt
(30:58) Do agents still need SDKs?
(31:44) Skills as documentation for agents
(35:13) Less code, even when agents write it
(36:44) Agentic development and the offshoring analogy
(37:56) Software factories and the plumbing underneath
(41:21) Keeping up while the standards change
(44:44) What Svix is launching with Poke
(47:08) Why Svix released the Poke spec early
(48:34) Building for your customers' customers
(50:08) Clerk, Resend and webhooks in the background
(52:49) Where to follow Tom

Links from the episode:
Connect with Tom Hacohen:
Connect with Chain of Thought host Conor Bronsdon:
More episodes: https://chainofthought.show

Thanks to Svix, presenting sponsor of season four of Chain of Thought. Svix delivers billions of reliable webhooks for startups and the Fortune 500. Get started at https://link.svix.com/cot. Qualified startups will get $12,000 in credits, $50,000 for YC companies.

Creators and Guests

Host
Conor Bronsdon
Creator and Host of the Chain of Thought Podcast
Guest
Tom Hacohen
Founder and CEO of Svix

What is Chain of Thought | AI Agents, Infrastructure & Engineering?

AI is reshaping infrastructure, strategy, and entire industries. Chain of Thought is the podcast where builders reason through what's changing. Host Conor Bronsdon sits down with the engineers and founders shipping AI in production to get past the hype into what's working and what isn't. Episodes cover model infrastructure, inference, agent frameworks, evaluation, and developer tools.

Guests have come from NVIDIA, Google DeepMind, AMD, Databricks, Vercel, and more. Every episode carries a full transcript and show notes at chainofthought.show. New episodes weekly.

Conor Bronsdon is an independent consultant and angel investor in AI infrastructure and developer tools. He led technical ecosystem at Modular, acquired by Qualcomm in 2026; led developer awareness at Galileo, acquired by Cisco; and ran developer marketing at LinearB, where he was GM of the Dev Interrupted podcast and community.

Views expressed by the host and guests are their own.

[0:00] Tom Hacohen:
JSON. is an abomination. So maybe webhooks are duct tape. Um they're just like good enough for the mirrors and they work. maybe we don't need SDKs anymore. I don't think you could have found two identical implementations of webhooks in the wild. when the laptop is there is is in your backpack and it's closed. Um that's that bridge also goes down. when you open the laptop you're gonna start like get a storm of retries from you know like 12 hours ago. Even with agentic flows, you can aspire to have simplicity.

[0:04] Conor Bronsdon:
Okay.

[0:32] Conor Bronsdon:
Welcome to Chain of Thought, I am your host Conor Bronsdon and today we're talking about webhooks. Webhooks Are the underpinnings for so much of the angentic internet today? And Svix delivers billions of webhooks for startups and the Fortune 500. their co-founder Tom Co-created the standard webhook spec that OpenAI. Anthropic and Google Gemini all adopted. Svix's founders have argued for simplicity since well before this AI wave. years at Samsung on the graphics tool kits. behind Tizen, then an encrypted sync side project. Or the webhook problem. Found Tom. With me today is Tom Hakoen, founder and CEO of Svix. Tom, lovely to see you. Thanks for joining me.

[1:16] Tom Hacohen:
No, thank you for having me. Looking forward to chatting.

[1:20] Conor Bronsdon:
Tom, why has the integration layer stayed this challenging for the last couple of decades. And How are you? trying to solve it now as it becomes so urgent for AI.

[1:33] Tom Hacohen:
In a way It's much, much easier than it used to be. I don't know if you remember those days, but now we have REST and JSON and HTTP, and back then it was custom TCP. um protocols. Everyone had their own custom thing. Um So I think things have gotten wa much, much better. But also, they're just way more integrations. I think it's like Jevon's paradox, right? It's kind of like the more. The the easier easier something becomes, like it's just You know? Just like Exactly. Gets used way more.

[1:58] Conor Bronsdon:
Yes. The more it gets used, yeah. And before we go too much farther, I do want to say a quick thank you to you and this fix team for sponsoring season four of Chain of Thought in this episode. You can find out much more about them at link.svix.com/slash COT. And of course, we're going to be talking all about webhooks and how Svix is approaching them today. Webhooks are viewed as a very unglamorous piece of infrastructure. And yet I think anyone who tries to build their own version at scale starts to realize the complications that come with it.

[2:02] Tom Hacohen:
It's just like with volume you get like different challenges and and more challenges. I think the main thing for that is that because People do things differently, right? So like when you have like one, 10 integrations, you have. people doing maybe 10 different things. Maybe less. But when you have a thousand of those, you're going to have a thousand different things. Yeah. And and it's something we actually we saw with with standard webhooks as well. Um you know like when we created it The understanding or the realization is that everyone Everyone was doing everything was doing things differently.

[2:31] Tom Hacohen:
Like, literally, I don't think you could have found two identical implementations of webhooks in the wild. And just by reducing that space, I think we made integrations easier again. So your question, I think they're Much easier, but also like much harder um and and the harder part is just really because of volume.

[3:23] Conor Bronsdon:
You found this problem. Building an end-to-end encrypted sync problem. Not by setting out to build webhook infrastructure in the first place. What did you hit that made you stop and say, hey, I'm going to build a company and a standard around this.

[3:32] Tom Hacohen:
Mm. Yeah. At my uh it's called EteSync, the previous company, and actually it's still running on the side. Um If anyone wants to use it. Can people kept on asking us for webhooks? And we kind of like analyzed it. You know, we didn't have a lot of resources. And we realized that it's gonna take us. A lot of time. And even if we did build it, it's gonna take us a lot of effort to maintain. And even if we did manage to find the time to do both of these, We it's You know, we built like such a great product and the webhook system is gonna be Like a lame duck in comparison, it's not gonna be great because we we can't we can't You know, we can't spend the 10x the amount of time to make a great weapon product.

[4:17] Tom Hacohen:
And I wish I was smart enough to realize at that point that there was a product there. But it was only like a a few months after this, um, that A friend of mine Ask her silly questions about webhooks on a Slack channel. And I started explaining to her: you should do this, you should do that, and you should do this, and don't forget about that. And and I realized with like all of this complexity that I was explaining that She couldn't care less about any of this.

[4:39] Tom Hacohen:
She just wanted to send webhooks and not have to think about it. so I asked her, like, Hey, if I build your product, will you pay me? And she said like, yeah. And then someone else chimed in and said, Yeah, we want webhooks as well. And kinda you know, like very low conviction at that at that point, but I I was like, you know what, I can build something for you. I built it for them and then I realized that actually it adds a lot of value.

[4:58] Tom Hacohen:
And there were more and we launched it on Hacking News and we got the perfect mix of You're wasting your life. This is such a silly idea. Um no but you know you need those as well because that's that's why not you know ha Not everyone built it already. You need to have the haters. And then obviously the other half, which is, Either We need webhooks now. Let's chat. Or This was a huge problem. Where were you a year ago when I had to build it myself? So it was like a great motivation boost, like the hack and news launch.

[5:05] Conor Bronsdon:
Yeah. So Let me be one of those lovely hacker news commenters. Some people consider webhooks a bad primitive that should be gotten rid of but folks can't figure out how to. What do you say to those people who are Mm, maybe the grumblers in the situation.

[5:32] Tom Hacohen:
Okay. Yes. Um that I agree with you all. You're 100% right. But I think it's like it it's an example of like worse is better. Um you know, like JSON. is an abomination. Um I really I I there's n n no not a lot of good things to say about JSON. Yet? It's so widely used. And it's, you know, it's ubiquitous. Um, and it's just so easy for people to use. It's like worse is better, and you get to use and. You get to add value by using JSON.

[6:13] Tom Hacohen:
That you wouldn't by using, I don't know, um Protobuf or whatever, you know, other thing that it may be better, maybe worse, or even message back, right? Which never caught on. Um so I think webhooks are like that. drive down like uh uh any downtown s in like in a major city, you'll see some like mirrors attached with duct tape. So maybe webhooks are duct tape. Um they're just like good enough for the mirrors and they work. So

[6:34] Conor Bronsdon:
Okay. All right, well if we're gonna hate on JSON, what other tools do you hate, Tom?

[6:42] Tom Hacohen:
Oh man.

[6:43] Conor Bronsdon:
Yeah.

[6:45] Tom Hacohen:
I think I have a love-hate relationship with everything in the world. Some things up. I think maybe Rust. Rust is the exception. Other than compile times I love Rust. Um but I think Yeah. Such amazing things, but the problem with developer tools and libraries and protocols and all of those is that it's impossible to iterate on them. J because I think JSON could have been much better if we were able to just like scrap it and continue to like iterate and iterate and iterate, but we just we can't, right?

[7:13] Tom Hacohen:
Because it has to be compatible. Um so I don't have a problem with Jason who ever created it as much as the frustration with the fact that we can't make it better. I think that applies to everything.

[7:21] Conor Bronsdon:
And one of the challenges to improving these primitives today is the amount of signal we're getting as agents begin to 10x the internet and 10x the number of calls that are being made across every service. as agents move to become the primary users of many apps. What is specifically breaking? when the thing receiving the event is now an agent. on a laptop or maybe in the cloud rather than a service in a data center.

[7:58] Tom Hacohen:
So Eight. Events are kinda unique in the sense that they're controlled by the sender, not the receiver. So like when you make an API call, you you know, like You can make an API call to API/slash v1, api/slash v2. You control, you know, like you can define the schema that you want. You can define a lot of, you know, like if you think of GraphQL, for example. And I think events have traditionally been way more strict in what they Um what they send. So I think that's like one of the challenges is that they don't They don't have the same um flexibility that we have in like other other places that agents enjoy so much.

[8:36] Tom Hacohen:
Um And the other thing And I'm not even talking about local and I'm gonna talk about like local consumption in a second. The other thing is as well is that like the way events are created, because they're like created asynchronously. They they kinda like they happen like almost outside of the session. So, you don't know when you're going to get an event. Are you going to get an event from a previous session? So, controlling those becomes a very complex problem. Um The last part, which is what you just mentioned is also like if you're receiving it on our laptop or like some you know, like some very ephemeral environment. Um The way to consume events at the moment just hasn't It it just doesn't support those use cases very well. Um

[9:14] Conor Bronsdon:
So Let's talk about that scenario where An agent's running locally. It's on a laptop. And The laptop gets close. We've seen this meme quite a bit lately where Uh folks are keeping their laptop open as they walk and We're seeing little 3D printed claws to keep your laptop uh open. Obviously, you can spin up agents in the cloud. This is more and more how agent fleets are built. Many things are running locally. What happens if that machine does get turned off? And Where are all these events going? Uh Things act say overnight.

[9:54] Tom Hacohen:
Yeah. So at the moment, the situation is not great. And we can talk about it later. We added some plug-ins for Hermes and OpenCloud to make it slightly better. But at the moment what people often do Is the user tool like Ngrok or, you know, Slick Splay or any tool that forwards like a Cloudflare tunnels, any tool that forwards to like a local server. The problem is that You know, when the laptop is there is is in your backpack and it's closed. Um that's that bridge also goes down.

[10:21] Tom Hacohen:
And those events just If you're you know, you're from a normal provider like Stripe or Svix or any of the Svix customers, they just retry and retry and retry and then they kind of fail. And and then like when you open the laptop you're gonna start like get a storm of retries from you know like 12 hours ago. that's kinda like the current state of the art or like the current state of things. And we We actually created I think it was like three years ago w way before agents, we created what we call like polling endpoints.

[10:49] Tom Hacohen:
Um which was essentially a way to to let people consume them. you know, by polling. Instead of by sending. And the nice thing about that is that they can continue from where they stopped. So when you open the laptop again, The agent could just go and continue from exactly where they they last got the the the last event that they got and like start pulling and pulling and pulling and get all the events that they missed. Um Yeah, so I think if you use one of those tools, it's a much better experience.

[11:17] Conor Bronsdon:
Let's talk about that Ermes and Open Claw. integrations Right now, because I think it's really interesting that instead of pushing to the agent. The agent in these integrations are pulling a spixed buffer over outbound HTTPS. And Essentially reversing the attritional approach here. It's also a great example. of Building for the agent scale of I think the oncoming, or maybe we're already there, agent economy. And we're seeing the success of these agents and they're widespread. with the personal agent competitors from Grockbot, from Muse, from Dots by OpenAI.

[12:03] Conor Bronsdon:
There might be another one out by the time you and I actually release this episode. Uh ERMI's just raised, what, $90 million at a $1.5 billion valuation. It's clear that people are Whether the developers or whether the individuals are going to have more and more agents running on their behalf. What does this mean for How the product surface around Webhox changes and How Are you enabling this? on Russian gigantic economy.

[12:37] Tom Hacohen:
So First of all, I want to say that yeah, I'm uh I'm a hardcore Linux user. I love using local software. But I do realize that the cloud has huge advantages. So I think what you just described, we we're gonna see even more of it. Like I don't think Open floor enemies are gonna are gonna be like self-hosted for the long run. Like they're gonna I mean I think maybe they already do have like a cloud offering, but if not they'll have one very soon and they'll essentially complete directly with town and muse and all of those kind of things.

[13:07] Tom Hacohen:
Um just because like people don't wanna run. software locally unfortunately. Um In terms of how to help those um tools I think we can see two different ways of approaching it. Either we're gonna have like very Mm-hmm. Structured integrations. So, for example, Yeah, I don't know, Muse may have an integ direct integration with Stripe. And it's kind of like one integration setup, you know, the old fashioned way. It's kind of like a ve you know, a third party to a vendor to a platform kind of thing.

[13:38] Tom Hacohen:
And that will just work the way it has always worked. I don't think it has anything to do with the agents. Um but then the You're going to have a lot of other ad hoc integrations. So, like wherever the new Stripe is going to be in a few years. Or even like off-platforms, you know, some of our customers like up and running. You know, we have like recent and agent mail that like provide emails that our customers. And I think with those You do have to have like a very different way of doing it.

[14:05] Tom Hacohen:
The agent needs to be able to subscribe and define what they want to subscribe to and define. You know, the timeframe of the subscription and probably like even more things that we haven't even considered yet about like scoping the sessions of callbacks. And there is, I think, a lot more to be done. And And it's very hard. to know before knowing how people things are going to be used. So we talked about open clone RMIs. What we did there was more A reactionary? uh thing. We just saw that people you know, like you go to the Open Cloud Docs and you see Um you see like Very complic complicated docs about how to set up webhooks, and we figured we can make this much easier by just having this like shim layer to make it like a polling thing. I know, does that answer your question?

[14:52] Conor Bronsdon:
I think it does. And It leads me down a few different tracks here. We had Dhruv Bhatra, one of the co-founders of Vutori, on this show a few months ago. And he argued essentially that Most of the web will never get APIs for agents, or at least next five, 10 years, because despite how cheap code is getting. organizations like let's say your local school The principal is probably never going to have a developer, or at least it's going to take several years to do it. uh redo the website for the school district.

[15:31] Conor Bronsdon:
and enable an API where your agents can now go schedule classes for your kit. And if that happens, it has to happen at a broader layer. There are so many compliance needs around this, government moves slowly. Th this is the long tail of the internet.

[15:48] Tom Hacohen:
Yeah.

[15:49] Conor Bronsdon:
This implies that Browser use is going to be very important for agents. Um I think we're seeing that already with some of the tasks we're asking Codex and others to take on. How does this long tail infrastructure build out? Impact the webhook layer and other primitives that you're thinking about around. this new agent operating experience where agents are the layer by which we interact with the internet, but Much of the internet is still built for The last era.

[16:22] Tom Hacohen:
Yeah. I I don't th Maybe it's a cop out, but I think we're kind of like almost exempt from this discussion in the sense that the most tech-forward. Businesses will have APIs and webhooks. So, even if the long tail is not there, like I think most agentic traffic, most agentic integrations, and most economic value. is going to be driven over APIs and webhooks. Um so yeah. the government may not adopt AI, but actually America.gov came out a few weeks ago. uh which is like a chatbot for the all-American uh thing. I I haven't tried it, but It just shows that even the government is adopting those t those tools.

[17:00] Conor Bronsdon:
Yeah, and I do think there is An opportunity because of what coding agents enable. To more rapidly adopt the infra for this technology than what we've ever seen before. And I'm very careful to see how that goes. Um, because we may be all be very off in our predictions as we have, I think, a hard takeoff that we're starting to experience. Though there is a lot of cleanup to follow that, and to your point. The economic side of it is very much going to be tightly tied with agents. I think that's becoming pretty clear.

[17:12] Tom Hacohen:
Yeah. Exactly. And you just actually reminded me of something. Um I think it was like fifteen years ago or something like that. I used to go to this uh sports class And they had the most terrible web UI. And I just wanted an app, you know, that would let me you know, favorite teachers and Look at the schedule in a much more sane way. And I just hack together like a s a You know, a small proxy that essentially parsed the website and then served JSON instead.

[17:57] Tom Hacohen:
Um and then like built an app on top of that. There is a chance that we're gonna see a lot of kids. you know, hacking around or building proxies around the schools. Or even Yeah, like Tech forward principles, like utilizing those tech kids, like to build something like that. for them.

[18:14] Conor Bronsdon:
I I'm going to be fascinated to see how that all plays out because It is so clear we are entering an era of builders and Anyone who has an idea can go do it as long as they can explain it in plain language. Or even ramble about it over a microphone like we're doing right now. So I do think it's a really exciting moment and Webhooks is one of those layers that actually enables a lot of that. Maybe undervalued by builders.

[18:18] Tom Hacohen:
Yeah. Exactly. Yeah. Um I mean we're trying to fix that. So hopefully by the next time we chat it's gonna be even better.

[18:49] Conor Bronsdon:
Yeah. And one of the ways that I know you're trying to fix that has been helping set standards for the industry. You know, I mentioned it. earlier but Standard webhooks has been adopted very broadly. Uh by A already. What was the value in setting that standard? And How has that driven

[20:02] Tom Hacohen:
Correct.

[20:18] Conor Bronsdon:
adoption from these major companies like OpenAI.

[20:24] Tom Hacohen:
Yeah. So They're not a customer, not yet anyway. Um but I think we created it for like two reasons and One is selfish and one was not. And but even the selfish one wasn't about affecting revenue. And so the main reason was that we just identified this huge gap, and everyone was doing everything differently, and it was just a huge pain. And we have this like We have a webhook ingestion tool called SvixIngest. And and we have we we have so many integrations we need to build new ones all the time because like still pe you know there are people that don't use Standard webhooks.

[21:01] Tom Hacohen:
And it's a constant it's been a constant pain. So we identified that pain and we just wanted to fix this. It was something that annoyed us, and we realized that we're well positioned to do that because. You know. A lot of our customers are going to adopt it. I mean, actually, all of them ended up, almost all of them ended up adopting it. Um So, it was like a very strong position. We also found. Well, while we were working on it, we found like great partners that also identified this problem.

[21:24] Tom Hacohen:
So, like, you know, like Kong is the API gateway, and they They never had a webhook verification. A module because Where do you even start? You're going to have to build a thousand of these or a hundred thousand of these. I don't even know how many different web centers there are. But now with standard webworks, they were able to. And same with ngrok. Ngrok now automatically Does webhook verification? They can do that because standard webhooks exist. So I think the first part was just that, like identifying.

[21:51] Tom Hacohen:
Um identifying this gap and just solving for it. The second part, and this is the selfish one. many cell scores we had You know, the team coming up with their own You know, random. as signature schemes and their own random requirements. Mm-hmm. Like half of the sales process was explaining to them why there are like security issues in this implementation that they chose, or like how to make it better. And oh, if you change that, you know, it's going to be better in that way. And it just was a huge drag.

[22:20] Tom Hacohen:
And we realized that It's not even that the ecosystem's gonna benefit, even like us. Specifically, we are going to benefit from just having like a standard way of doing it that we could just refer to and just like, hey, listen, this is sanctioned. by literally everyone at the moment. Like everyone is using it. You can either be part of that ecosystem and benefit. Some of those tools that are being built, you know, we have libraries that we've built for like 12 different languages that it can just use.

[22:44] Tom Hacohen:
Um like unbranded, just branded with standard webhooks. And and and that was a huge unlock as well for us in that sense. Um I don't remember if we thought about it ahead of time. Or if that was like the spark of the idea, but it's definitely been like a selfish win for us.

[23:00] Conor Bronsdon:
This will give away to our listeners when we recorded this, but I see that. Google Gemini has actually just announced finally, their own personal assistant to compete in this space. as well just a few minutes ago. So I think that speaks to your vision of the future, where Everyone has more and more agents acting on their behalf. And it speaks to the value of these permits we're talking about. the need to be excellent in these areas.

[23:30] Tom Hacohen:
Yeah. And I think You know, like if you consider, we talked earlier in this call, like we talked about like the number of integrations. Um so now we have what like seven eight major um like agent Hart like I don't know what to call them, like agents, I guess. And they're gonna be even more. And those each like has to integrate like I don't know, like a thousand different services. So Just another example of like the explosion of different integrations that you all have to do. Yeah.

[24:01] Conor Bronsdon:
So You know, we've talked about some of those big names: OpenAI, Anthropic, Ludamin AI, they've all adopted Sander Bubhooks. It's fueling their Agent Layer. I would assume that Anthropic is going to come out with their own personal agent soon. That's the rumors that's coming out before their IPO to really continue to add to their accomplished in the space, though with the recent news that Opus 5.5. can be and is currently being used to power Um Grokbot, I I do wonder if they're going to take a a partner play instead here.

[24:34] Conor Bronsdon:
Can you tell me a bit more about the actual steering and establishment of the standard here? And Uh why standard web hooks is able to Begin to shape where you see the industry go in the future. Because I'm sure there are some folks listening who are either. mildly skeptical. They don't know anything about the standard. Or maybe interested, hey, I want to get involved, I want to try building on the standard. Uh what's the governance approach and How are you thinking about this spec as you move from A one point oh version to two point zero.

[25:08] Tom Hacohen:
Yeah. Um so first of all we have A committee? Um that committee um Besides like on what what gets done. And we we speak in public. You know, other than um Other than like small like notes to each other, or like to celebrate small things, you know, like When when I dunno, like one of the labs adopted standard web hooks, we kind of like sent a message to each other. But other than those, everything is done in public. I think like every spec There is you know, people that are more proactive and active.

[25:40] Tom Hacohen:
like in the committee and and some that are less, but like everyone has been weighing in on all the important stuff. Uh since the beginning. And for people on you know, that want to contribute. Yeah it's all on GitHub. So you can change the website. You can you know I think like Mercury recently um We released like an Elixir library that's compatible with standard web hooks, so we have that linked there as well. Um yeah, you can just c We can just do things, I guess. Um really just like go to our GitHub and just contribute.

[26:08] Conor Bronsdon:
And we'll certainly link that GitHub in the show notes. And I have made a couple of very small docs contributions myself and need to dig. More in depth. It is really cool to see how the industries come together here. I know you have. team members from Zapier, Twilio, Sukbase and a variety of other great companies on that steering committee with you. Um How are you? Going to continue to evolve the roadmap here? Like, what's the plan that the committee currently has to ensure the standard becomes more and more valuable? So that You can uh head off some of the criticism around you know retry guidance or anything else that is coming through.

[26:50] Tom Hacohen:
Yeah. So one approach that we've taken with standard webhooks that I think is different to how standards are usually made. Um is that we we try to follow real world usage. Instead of Lead it. So, like, I think a lot of standards kind of like they sit in a lab, they try to define something, and then they write it out, and then they adopt, you know, either hope people adopt, or maybe if it's something that was written by Google, Chrome just adopts it and kind of like powers through.

[27:15] Tom Hacohen:
It just gets it, you know, in everyone's hands. Um and here we just we see what are the gaps that people identify, what people need to do, what people have already been doing and it's been working, and we just we we do that. And that that doesn't mean that we wait ten years after. We're not like um You know, the C standard is famous for like adding things like twenty years after people have been leading them, like we try to be a fast follower. And like one example is Um Mm.

[27:41] Tom Hacohen:
MCP is adding WebHoc support finally. Um And We've been chatting with them about it and they're going to use standard webhooks. And one of the One of the things that they wanted, they wanted what's called CRC. But essentially what it means is it's a way for the sender to verify. Yeah, that the client actually knows the secret. So I can verify that the client is actually trusted. Um I don't think that's necessarily a requirement for most WebHooks implementations, but I w Yeah, some people Definitely have CRC.

[28:14] Tom Hacohen:
And that's an example of something we identified that people need and we're going to probably add to the spec in the very near future in order to support. Um you know, in order to support them. So it's kind of like, again, the fast follower can be even pre-launch. Like we can follow people as they're speccing things out, but we still try to see. what you know, what people need and use and kind of like help them there.

[28:34] Conor Bronsdon:
So if we think about your APIs as the new reader for models. And models are now writing integration code for pretty much everything we do. Yeah. I think any company coming out of Silicon Valley is true and any company is kind of on the bleeding edge and we're going to see that. continue to be more and more true as the Agent tech economy and digital coding expand and eat. other types of software engineering. If models are now writing most integration code and will in the future write All of it. What changes about how we design and document. are different surfaces.

[29:17] Tom Hacohen:
Yeah. Mm-hmm. And it's kind of like tied to what we talked about earlier in this call. One thing I'm excited about when it comes to agent integrations. Is that Agents are way more adaptable. So you can actually iterate on APIs like fairly quickly. And kind of like to call back on what I said, my biggest frustration with JSON was that you're not able to iterate, you're not able to make it better. You know, there's a company by a friend of mine. It's called like Restless. And what they're trying to do is to imagine.

[29:47] Tom Hacohen:
Reimagine rest in the world where it's just an agent consuming it and it can change all the time. And I think that that's like a very interesting approach. And and we see it as well, you know, like with the CLIs. They don't need to have like a set, or even with MCPs, MCPs don't have like a set-defined API that will never change. The tools can change on every invocation. And the agent is going to be fine with that. So I think that's like one very interesting thing.

[30:10] Tom Hacohen:
Um the o other thing that I think about every now and then is S D K's. SDKs were the most important thing for the longest while. And I think a lot of people were saying how SDKs are going to be essential for agents, and maybe that's true, I don't know. Um But I also I feel like in a world where the agent can actually just like hack together on a you know, an SDK. It's um Maybe An SDK is not that necessary. I mean, we still we have some like magic in our SDKs that like we wish Mm-hmm.

[30:37] Tom Hacohen:
everyone would use um Like, you know, like um retri you know, re automatic retries and and and and a few other things that I don't think the agent will necessarily hack together when they do it, or at least not hack it in the way that's compatible with how we do it. Um but other than that I I I don't know, maybe we don't need SDKs anymore. No idea. Play agents. Yeah.

[30:57] Conor Bronsdon:
All right. Let's go down this rabbit hole.

[30:59] Tom Hacohen:
Okay, let's do it.

[31:01] Conor Bronsdon:
We don't need SDKs anymore. What does that mean? For builders, how should they be adapting their perspective?

[31:08] Tom Hacohen:
Yeah, so Caveat. for agents. I think also like Humans still benefit from um from SDKs and I think SDKs do have Value. Like, I wish the agents used them, but like, I don't think they're a blocker anymore. And how people should adopt. I think people are adopting already. They're kinda they're building MCPs. Uh which is essentially You know, like it's an S D K less. integration that is like very dynamic. At the beginning it was all also like Maybe it still doesn't have like response schemas. I don't even remember. But it was like it's very dynamic, um and I think that's that's how people are adopting adapting, sorry.

[31:44] Conor Bronsdon:
What are the other changes you're anticipating? skills is an obvious area where we suddenly have this new. layer that we are leveraging with agents that was I guess you would s call it you know, tracked business law. Previously, and it's a way of adding some semi-determinism to these non-deterministic workflows that agents are driving. Yeah. Svix has ship skills rather than more docks in many cases. because skills are almost an active form of documentation. But there's obviously a ton of updates that come with this and Documentation goes stale while types don't, as I think you mentioned back a couple of years ago. How are you going to keep skills from rotting the way? per framework recipes did in the kind of lang chain era.

[32:33] Tom Hacohen:
Yeah. Um I wish I could sit on a high horse and tell you that we have some solution, but actually, I think they've already gone stale at some point. We had to fix it. Um But definitely like You know, one area that we're slightly behind, I think a lot of people are doing very well is like, um Um evals. And I think the nice thing about having agents as your customers. Is that you can actually ask them to work for you 24-7 and kind of like try things out and see if they work, if things are better, things are harder.

[33:09] Tom Hacohen:
And so that's definitely one way of doing that. Mm-hmm. you know, more and more. um companies are doing and The other one is that we try and it's kind of funny, you know, like a lot of the stuff about Igentic Feels both profound, but also exactly the same thing we've been doing for like 30, 40, whatever, whatever long years. And so, like, one principle is, you know, keep it simple, silly, and like, oh, dry, or don't repeat yourself. the docks are slightly different or like cover different ground.

[33:39] Tom Hacohen:
Then the you know, API reference docs. And then again, compared to like the skills. So the skills, they reference the other two. And they actually have a lot of Just original content that we don't have in the other places. Like we use the skills more as almost like a solutions engineer. It's just like, hey, you know, like, if Conor is trying to integrate, ask them whether they want to go this route or this route or check their code and verify that this is the way that they do things, which I don't think you can do this, like, choose your own adventure.

[34:07] Tom Hacohen:
In docs. But I think sk or or or even if you try to do it, I don't think it's going to be very useful for humans, but I think for agents it's like very good. So that's why the skills are very agent orientated. And then the docs are you know, what they've always been. Handwritten for humans.

[34:23] Conor Bronsdon:
Yeah. One of the ways I've personally adapted to this is I ended up building A skill. To keep my docs more up to date and turn it into a GitHub action that I run. On my larger repos and have an agent periodically run in a loop. It's just called SSOT check. Single state of truth check because I have been noticing that drift is so easy, especially when agents are just continually updating docs, adding comments, et cetera. And so. Having it as a deterministic workflow that runs.

[34:53] Conor Bronsdon:
As part of GitHub Actions, and then is also run by an agent regularly as a double check, has helped with this. It's uh a fight against the continual iteration and the mass amount of code we are Creating in this current era. Um And I I don't think that's a solved problem by any means.

[35:13] Tom Hacohen:
And I think like previous years as well. We should aspire to have less code. even if the agent is creating it. Mek I mean first of all for like For reasons of current limitations of context windows and all of that. But also, just simplicity is the elegance that comes with simplicity. Just means that agents are less likely to make mistakes as well. Because I feel like a lot of the agentic development flow that we have nowadays. It's kind of like, okay, we're going to make a mess.

[35:39] Tom Hacohen:
But we don't worry, we have these tools to review BPRs that's going to catch some of the mistakes and we're going to fix those. And then we're going to find we have an agent that monitors data docs, so we're going to catch those and quickly fix those. And then we end up with a huge amount of patches and top full patches and just more complexity instead of just like taking a step back and simplifying it. And I'm not advocating for Mm-hmm. everyone taking a step back and writing it themselves, but I think Even with agentic flows, you can aspire to have simplicity.

[36:05] Tom Hacohen:
And I think I really wish that it doesn't get thrown out of the window with all the software factories or. high ut utilization of agents.

[36:15] Conor Bronsdon:
I do think this is a challenge because Often our best software engineers are the ones who spend the time to figure out how to simplify and remove code. And yeah, we can. Do that with agents where you assign agents to think about potential work on this. But there's so many other agents that are already building a ton more code that it just becomes a a losing game, it feels like.

[36:36] Tom Hacohen:
Yeah. Yeah. Yeah. Yeah. I mean, I think y so I I spoke with a friend of mine the other day and I think he said something. That was interesting. I don't know if he's right. I don't think he's necessarily right, but it was still an interesting thought. And he said that The way people approach you know, approach hygienic uh development at the moment is very similar to how people um approach offshoring. Like, what a 10 years ago, 15 years ago, which is like, oh, oh my God, you know, like an engineer cure cost.

[37:06] Tom Hacohen:
100 bucks, whatever, an hour, whatever the time you did, there, it's only like 10 bucks. So we actually can have like way more engineers, way more code generated. And we're just gonna throw more, you know, more. intelligence as like humans versus uh here or the AI. At the problem and just like generate more and more and more. And he was arguing that this is gonna reverse slightly because of that. And I don't know if that's true, because I think here the the scale is just like infinitely larger.

[37:30] Tom Hacohen:
Like you can now throw a thousand or like a hundred thousand people at the same problem compared to like previously, which was like maybe 10x the amount of intelligence. Um But it's still like an interesting thought that like maybe Maybe technical debt will remain a thing. Maybe complexity will remain a thing. Maybe those things actually will matter. Um And then we're gonna have to figure out what are the techniques to like rein some of it in. Um and make it sane. Oh. Yeah.

[37:56] Conor Bronsdon:
It's a really interesting thought experiment. Because We'd We do have these tools now that can be Always on helping solve our technical debt. As it is created And yet If we're not careful, it's so easy for the continual development they are doing to just scale this problem. And I think this is why everyone's talking about software factories, right? Like, how do I Have software best practices. And maybe we'll get to a golden software factory. Maybe someone else already has, and I'm just out of the loop on it.

[38:33] Conor Bronsdon:
It is so hard to have it be continually solved, especially because, despite us adding many deterministic workflows as part of these. software forages, software factories. It's still non-deterministic. Models that are driving this, and so it's easy for there to be outflows and changes. And the space is changing so rapidly. That As the intelligence of the models grows. I think we're rewriting. everything regularly. I know it's certainly true in my personal projects where I get a new level of intelligence, let's say you know Opus 5.5 big step up.

[39:04] Tom Hacohen:
Yep. Mm-hmm. Yeah. And I think it's yeah, exactly that. Because like the even the amount of logs that are being generated, right? I mean, I'm sure a lot of it is just the fact that we have, you know, like a million agents running in parallel. But part of it is just that we're just wasteful because it doesn't matter and like we don't care about it and we have like solutions, but Even in our lifetime, I think. What, my first computer probably had like a I don't know, five megabyte hard drive or something like that, six megabyte, or something like of that order of magnitude.

[39:12] Conor Bronsdon:
in in capacity and Ability to just do long-running tasks. And I said, Hey, Opus, I need you to go review. all my key repos and refactor. And sure, it did that. It cut some lines of code, added some others, had some great suggestions to simplify. But it didn't lower the lines of code of the over repo that significantly. Unless you're being very intentional about like Okay, I want to simplify this down to the very core primitive, especially because of how much documentation we're creating. uh tracking of memories, tracking of context.

[39:46] Conor Bronsdon:
We're trying to get these agents more information. And that does just create uh a challenge around How much to provide and adjust I think it just speaks to the fact that we are we've changed the abstraction layer that humans are operating on. And It doesn't mean we've solved the layer underneath. There in fact may be some of the same problems. We're just changing. how we're experiencing and interacting with that layer. And The acceleration of code into all this plumbing that is already somewhat broken. Uh and maybe very broken in many organizations. may actually exacerbate some of these problems if we're not careful.

[40:52] Tom Hacohen:
Um And Know, like it's changed so dramatically, and I think we learned. We build new tools, we learn to adapt, we We have new techniques. And I wonder. You know, we're just like rerunning so quickly. I wonder if we will find I I think we'll f I'm sure we will find like good ways to like You know, wrangle all of this data and actually make sense of it without having to resort to, as you said, just outsourcing it and reading the synthesis. Um like maybe more deterministic ways.

[41:21] Conor Bronsdon:
Yeah, I mean we're seeing different approaches. This doesn't necessarily speak to the deterministic piece. But I know developers who have essentially an automatic architectural animation that runs in parallel to their IDE. Uh where they and this is something that's being added to things like T3 code and and other Agenzek IDEs. And You're seeing just a continual animation of what is the architecture of the system being built, where is data going. So you can go, oh, wait a second, this isn't attached to anything anymore. This is now vestigial.

[41:38] Tom Hacohen:
Interesting. Okay. I think So I I can tell you I don't know how everyone should be approaching it. I can tell you how I've been approaching it. It's been so hard to to try to guess how things may look like. So I think I found the best approach for for me and for us has been to just like keep our fingers on the pulse and continue to adapt as things adapt. And I mean, you know, try to try to see what people do at the periphery because some of what they do is gonna become the next step.

[41:55] Conor Bronsdon:
Um And I think that kind of visualization layer is Is interesting. I don't know that it's going to be the right layer of abstraction. as we go forward. But We're seeing this all change very rapidly. And Obviously, MCP, which you brought up earlier, that they are aligning with your work, which I think is fantastic. But they have changed so much since the launch of uh the the protocol. They went stateless in July. and move tasks out of the core protocol into an official extension. Um It's been crazy to see the development of some of these now widespread standards for the AI era.

[42:35] Conor Bronsdon:
Let alone things for the human-led era that are still being pulled forward into the Adventic era. I would just love your take on how we should all be thinking about navigating this next layer.

[43:24] Tom Hacohen:
But I think as long as you like you stay And you keep on track the frontier, you're gonna be Well ahead of the curve. Um just because like you you see what's coming. 'Cause I think that's the best we can do. Like, I I don't think Ye you know, like coding agents were not that good. like ten months ago. Um an agent's open claw was not a thing 10 months ago, right? Um or or whatever, maybe twelve once or something like that. So like MCP was not a thing like a year and a half ago, I think, I don't know, two maybe two years at this point. So all of these Yeah.

[43:54] Conor Bronsdon:
Uh God, yeah, almost two years, isn't it? 'Cause that was what, November 2024 or September? I forget.

[43:57] Tom Hacohen:
Oh, see ya. Yeah, I think something like that. And And anyway, we're like we're now at MCP 2.0, right? Like essentially, like the going status that you mentioned. Um see if you try If he tried to get but but So, if you try to guess like where things are gonna go, like you haven't, yeah, there's there's a zero chance to to know. But like if you were you know, lurking in the In the MCP discussion boards, you kinda like you, you know, you checked out their GitHub, you'd see that actually they work in the open and stateless was coming.

[44:28] Tom Hacohen:
For like a almost a year or however long it took. Um So so I think as long as you just like you stay informed Yep. and just try to keep up, I think that's that's the best approach. Yeah, I don't think anyone any of us can predict what's coming next.

[44:44] Conor Bronsdon:
Well speaking of what's coming next. I know Sphinx. is making some big moves. What's new for you and your team?

[44:53] Tom Hacohen:
Yeah, so we we just launched something Um new it's we call it poke. Um so it's like publish once uh know everywhere. Oh. Now it's where? Kickoff everywhere? I don't remember. It's a new acronym. But essentially, We were just trying to think of like how webhooks and how events and especially events between different services. should evolve in this new era. And again, I say we're trying to think. It wasn't like us sitting in a room somewhere. It was just like seeing how our customers actually use the product, seeing how agents use the products, seeing all the challenges that we face.

[45:25] Tom Hacohen:
Again, I keep on see using the word agent, but as I said, a lot of this stuff actually like Predates agents. Like a lot of the best practices that you should apply here. have always been best practices. Um That's a very long way of saying, though, that we're just like trying to do event-driven for agents and event-driven to make agents to make it easier for agents to subscribe to events in different services and to also trigger asynchronous events, because all of this. It's kinda like the the parallel that I made earlier to offshoring.

[45:52] Tom Hacohen:
I think like it also applies here in the sense that like you ask someone to do it and they get back to you in like four hours. Instead of just like the normal API response time of like 100 milliseconds or 10 milliseconds or whatever it is for your service. Um so I think some things have inherently changed. And also the The scale of our ambitions of what an A what an API should Do? has also increased. Um Yeah, now it's very normal to ask an API to generate a full feature.

[46:17] Tom Hacohen:
You know, maybe not a full feature, but like a Seinfold episode or something like that. That's like normal, right? Um So We want to make sure that You know, we follow that worse is better of JSON that I mentioned earlier, or REST, or maybe just like the simple is better, is maybe a better way of putting it. I just make it easy for people to do the right thing from the get-go. Instead of someone having to come you know, five, ten years from now with like a standard web-esque solution and kind of like try to rein in the complexity.

[46:49] Tom Hacohen:
I'm a big fan of simplicity. Huge fan of simplicity actually. So Trying to help Keep things simple, I guess. That's maybe the w right way of putting it.

[46:58] Conor Bronsdon:
Perhaps the most valuable skill in the current era is Figuring out how to create simplicity in certain areas. I I love that you're taking that perspective. Tell us a bit more about Poke and what's happening with the Poke launch. Who is using Poke off the bat? Where is it? being valuable for teams.

[47:05] Tom Hacohen:
Yeah. Oh man, it's it's very, very early. Like we don't even have the the product full fleshed out It was more just we wanted to release the um Mm, release the spec. um out in order to to get people to like give more feedback. Um early on so it's while it's still malleable. Mm-hmm. And we do You know, it it du it does rely on the On the tools that we've already built for our customers. So, for example, one of our customers is agent mail, and they do agent identity and email.

[47:44] Tom Hacohen:
Yeah. you know, we build tools for their agents, which is like a lot of open clusters and OMEs and all of that to kinda like Very easily subscribed to notifications about when an email is received and all of those kinds of things. And just make it very easy for them to To react to that. So, like, it's, you know, it's kind of inspired by that work, inspired by the work we did with polling endpoints, and inspired by the work that we did with standard webhooks, and I mean, and uses some of these things as well.

[48:08] Tom Hacohen:
And it's really just It gives us the opportunity as well to sit back and kind of like drop the bad ideas. and double down on the good ideas.

[48:17] Conor Bronsdon:
And obviously, if you want to learn more about everything that Svix is building, You know, we have our own link for the podcast, link.svix.com/slash

[48:17] Tom Hacohen:
'Kay.

[48:26] Conor Bronsdon:
C-O-T. So you can learn much more about Poke, how they're approaching standard webhooks. And there's more info on standard webhooks as well at standardwebhooks.com. I would love to Close us out here. Tom Bye. Asking you Where Our web hooks and the work you're doing at Svix, providing the most value for customers today. What are you hearing from teams you work with? And What's your dream for what you want to provide them over the next year so that they are enabled in this new era?

[49:02] Tom Hacohen:
I think, you know, being like a a middleware I guess, and I mean I mean it in a nice way, I guess, about us, but like we you know, we sit between our customers and their customers. We we add value both to our customers and I think and their customers as well. I always love hearing about how we add value to their customers, even though we're here to serve our customers mainly. But The fact that people can build integrations that they can rely on. And you know, like, and have just a very easy and great experience.

[49:30] Tom Hacohen:
And, kind of, like what the example I gave earlier about digentic flows, as well, just like everything just works. Always makes me very happy. Because I think But th They just like normally underserve. Like, you know, like if you ever use webhooks, you know how the terrible experience. It can be with some providers. And like by helping our customers Give their customers the best experience. I think their customers are getting like a huge, huge difference. For our customers as well, they stay up, they don't have to think about webhooks, they don't get paged, and they can be proud of what they've built. And but I think I'm really proud of what we help with, you know, with their customers.

[50:08] Conor Bronsdon:
What teams are getting the most value out of Svix today?

[50:11] Tom Hacohen:
Um a lot of names that I can't say, unfortunately, but like names household household names that you all n would know. Um, but there are like a lot of like some of the best-growing startups out there. So, you know, like Clerk, which I think. If you've ever built with an agent, you probably use Clerk, and you probably use them as well. If you haven't built with an agent. Um You know, we helped them I don't remember what the original estimate was, but like we we helped them like shave months off their uh roadmap.

[50:39] Tom Hacohen:
Mm-hmm. They've just scaled and scaled and scaled and scaled and like never had to think about webhooks ever again, like literally. They just it just kind of like scaled in the background. Which is kind of like a funny place for us. I always kinda like enjoy and also maybe regret in some ways is that our customers forget that they're a customer. You know, because everything just works in the background. Um I mean they don't really forget they know if you ask them, but they kinda like they don't even think about It's not like a dashboard that you log in like, you know, five times a day because everything just works.

[51:06] Tom Hacohen:
Um Another two, if we're talking about like fast-growing startups, like resend, another one, if you ever built with an agent, you probably use resend. Yeah, they they've been doing been doing great and I think they've been leading the way in so many Areas like DX before and now with agent experience and GEO and all of those things. Yeah, so they've again been growing and growing and growing and I don't think they ever had to think about webhooks ever since they integrated it in the beginning. Yeah, I think those I mean, we've you know, we have like very large customers as well, I just mentioned like fast-growing startups, like large organizations like Build Time require.

[51:13] Conor Bronsdon:
Huge fan of Resen. I think one of the really cool things, and I am a shell here. Obviously, Svix is sponsoring the season, but I'm a huge fan of what the product is doing. And One of the coolest things is that you're also making it easy for earlier stage startups to get started. You're offering $12,000 in credits for qualified startups. $50,000 in credits for YC startups. And It just takes so much of the headache off of the plate of many of these founders and founding engineers.

[51:40] Tom Hacohen:
even more time and they also often come with like enterprise requirements that they will just have to build themselves and it's just like a huge slog. Um.

[52:11] Tom Hacohen:
Yeah. Um, I guess TwitterX is probably the best uh place, so like At Tom McCohen. Um very simple. Managed to get the handle. Yeah.

[52:18] Conor Bronsdon:
So. If you are someone who's listening to this episode, And Are you trying to level up your agent pronobes and your agentic engineering? Check out specs. Seriously, link.svix.com/slash COT. It is a great product, and you can get started for free and decide if you like it or not. Tom, thank you so much for spending the time with us today. It has been fantastic chatting with you. I appreciate you getting down into the nitty-gritty. uh the decisions around the next premonubes for the Aegentic era. Where can folks go to follow you and your work?

[53:00] Conor Bronsdon:
Fantastic. Well, Tom, it's been great chatting with you. Really appreciate you taking the time. And anyone who listened all the way to the end. Make sure you like and subscribe. Drop a comment if you enjoyed the conversation, or let us know why we're idiots about something we said. We would love to hear from it. Uh so thanks so much for listening everyone and Tom. Thanks again.

[53:18] Tom Hacohen:
Oh, thank you. So