DEV

Deploying AI agents is the easy part — governing what they do in production is where most teams come up short. This episode breaks down the three control-plane components every serious AI deployment needs: approval queues, immutable audit logs, and role-based permissions.

Show Notes

The promise of autonomous AI agents running finance, marketing, and operations sounds compelling — until something goes wrong and no one can prove what the agent did, who authorized it, or whether the right controls were ever in place. This episode of DEV.co cuts through the governance theory and gets into the practical architecture that separates a production-ready agent deployment from one that's quietly waiting to fail. Drawing on the full companion article on AI agents in production: approvals, audits, and RBAC, the episode gives technical teams concrete schemas and patterns they can actually ship.

Here's what the episode covers across three core control-plane components:

  • Approval queues as a data model: The minimum schema every pending agent action needs — request ID, correlation ID, risk tier, reviewer options, expiration timestamp, and a null-until-decided decision field — and why stale approvals should fail closed, not execute.
  • Action tiering and dual control: How to classify agent actions from auto-execute reads all the way up to two-approver gates for payments, credential grants, and production config changes — and why that dual-control pattern mirrors where regulators are heading for high-risk AI systems.
  • Approval fatigue is a real threat: Research showing that human reviewers approve over 90% of prompts (not because the actions are safe) and catch only ~14% of harmful ones, compared to ~90% for an automated classifier — and what that means for how to design the human-in-the-loop layer.
  • Immutable audit logs vs. application logs: Why these are fundamentally different stores, how hash-chained Postgres tables, Kafka-backed WORM storage, and managed cloud append-only stores all achieve tamper-detectability, and what each log entry must contain (including content hashes instead of raw sensitive values).
  • Retention and replay requirements: How PCI DSS, HIPAA, and other regulated workloads set minimum retention windows — and why the log must support reconstructing the agent's state at decision time, not just recording the tool call.
  • RBAC with agents as principals: A four-layer permission model covering agent roles (scoped per function), reviewer roles (with separate dual-control identities), operator roles (pause/resume/stop), and read-only auditor roles — all relevant to teams building with custom AI agent development frameworks.

For teams operating in regulated industries or scaling cybersecurity-conscious infrastructure, the control plane isn't optional — it's what makes the difference between an agent that works and one you can trust. Listeners interested in adjacent security and compliance concerns in AI systems will also want to check out the episode Secure AI Agents for ERP, CRM, and Finance Systems.

DEV.co

RFP.co

What is DEV?

Software and web development from the side that has to ship it and then live with it. Architecture decisions with a cost attached, scoping, technical debt, hiring and vendor selection, and the AI tooling question every engineering team is now answering whether they planned to or not.

Each episode takes one decision — rewrite or refactor, framework choice, build versus buy, how to scope a fixed-bid project honestly — and works through the tradeoffs, including the ones that only show up in year two. Written for engineering leads, technical founders and the people who fund them. Five or six minutes, no hand-waving.

Topics include rewrite versus refactor, build versus buy, scoping fixed-bid work honestly, technical debt you should keep, framework and platform choices, hiring and vendor selection, code review culture, and where AI tooling actually helps.

Produced by DEV.co, web and software development. Full details, services and further reading at https://dev.co