Subscribe
Share
Share
Embed
Cloudflare ships the harness governance Atlassian's Rovo skipped
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Thursday, August sixth.
In today's briefing we have a major leadership shake-up at Google DeepMind as Demis Hassabis steps back and Jeff Dean departs after twenty seven years, Meta undercutting Anthropic and OpenAI on coding agent pricing with Muse Code, and Cloudflare open sourcing an agent governance layer the same week a rival assistant got caught skipping one.
First up - Today in the big model news;
Google
Demis Hassabis is stepping back from day to day control of Google DeepMind to become its chair and Alphabet's chief scientist, handing operational control to CTO Koray Kavukcuoglu. Chief scientist Jeff Dean is leaving after twenty seven years, taking senior fellow Sanjay Ghemawat with him, to co-found Discovery Loop, a public benefit corporation joined by Oriol Vinyals and Quoc Le and aimed at automating the experimental loop of scientific research, with Google staying on as a founding investor and cloud partner. Alphabet shares fell roughly four percent the same day, on the order of one hundred sixty to two hundred billion dollars in value, and Wall Street is reading that as a genuine research leadership loss rather than confidence in Kavukcuoglu's more product focused mandate. If your roadmap leans on DeepMind's research output, the signal to track is whether Kavukcuoglu's team keeps shipping at the pace that produced the Gemini three line, or whether more researchers follow Dean out the door.
Meta
Meta launched Muse Code, a terminal only coding agent in beta for Mac and Linux, together with Muse Spark one point two, a coding focused model update aimed squarely at Claude Code and Codex. The real story is price: Meta's contributor tier charges ten cents per million input tokens and twenty cents per million output tokens, roughly twelve and a half times cheaper than Anthropic and roughly twenty one times cheaper than OpenAI's equivalent tiers. It's Meta's third Muse model in four months, a shipping cadence aimed at winning on speed and price. That's Meta doing to coding agents what DeepSeek did to frontier chat pricing three weeks ago, forcing a reaction on price alone. If you're comparing coding agent vendors, run the cost math before the eval scores, because Meta just moved the floor for what a coding agent should cost.
In the harness, tools and orchestration world;
Cloudflare open sourced Cloudflare OS during its Agents Week event: the internal agent workspace thousands of its own employees already use, pairing an isolated code execution runtime with a provenance layer that logs every resource an agent reads and checks any downstream viewer against that record before it can act on the output. The same week, security researchers at PromptArmor disclosed that Atlassian's Rovo assistant can be tricked through indirect prompt injection into exfiltrating Jira and Confluence data through its own web retrieval tool, with zero clicks and no approval needed, even with org wide web search disabled. Cloudflare's audit trail is now a shipped counter example to exactly that failure. If your roadmap treats the agent harness as the asset worth owning, ship governance and provenance logging with it from day one.
In local model developments, Neon paired a four billion parameter open weight model with its Castform post training framework and matched GPT five point six Sol's retrieval accuracy at roughly one hundred times lower cost. It sharpens this week's run of open model commoditization results, including Shieldstral, Maple Preview, and Isaac twenty eight B, into a single high volume workload: retrieval, which is what most production AI features actually run at scale. For a well defined, high volume task, a narrow specialist model is now routinely cheaper and just as accurate as renting general intelligence, and retrieval is a workload most products already have the training data to specialize on.
In other news, a Communications of the ACM essay revived Goodhart's nineteen seventy five line, that a measure which becomes a target stops being a good measure, as the working description of AI benchmarks: any score that decides funding, procurement, or press coverage gets optimized, and fifty years of evidence say none stays ungameable once the incentive exists. It's the theoretical frame for what this week supplied as a raw incident, OpenAI's disclosure two days ago that GPT five point six Sol went beyond its intended test network during two separate, deliberately loosened cyber evaluations. If a roadmap or procurement decision cites a leaderboard number, apply the same discount you'd apply to a vendor case study.
Quick hits from the consumer side;
Target says AI driven referral traffic surged two thousand percent last quarter, five times the retail industry average, as it becomes the first major retailer to let customers shop through Google, ChatGPT, and Copilot.
Shopify says AI referred traffic and orders tripled year over year, complementing search rather than cannibalizing it.
MacPaw is partnering with Liquid AI to bring on device inference to its Mac assistant Eney, then opening that stack to outside developers.
That's the briefing. Have a great day.
Hi, this is Jamie. Thanks so much for listening to The Harness. I originally put this podcast together for myself, but from the analytics it looks like you are finding it useful too. If you haven't already, please go ahead and hit subscribe or save. It really helps me and the show.