An AI agent deletes a payments safeguard that existed because of a disaster. The pull request is green. The reviewer is junior. Nobody ever decided the agent was allowed to touch that service. It just could. The agent acted. Who answered?

The framework
Scope the authority. Keep the trace. Right-size the human checkpoints to blast radius. Give the agent memory. Red-team the whole thing. The agent acts, you answer.

Inside this episode

• Intake — Once agents act and chain decisions across your systems, "is this model accurate" stops being enough. The new question is what an agent is allowed to do, and who's accountable when it acts.
• Flow — The five-move control model: capability envelopes enforced at the infrastructure level, the trace, tiered human checkpoints, ADRs as the agent's memory of why, and governance red teams.
• Outcome — A model program runs the five moves in a quarter: ADR coverage 35% to 78%, agent-authored drift from ~62% to single digits, and the trace catches a near-miss before it ships.

Also inside
Gartner expects agents in 40% of enterprise apps by the end of 2026 and 150,000+ per Fortune 500 by 2028; non-human identities already outnumber humans roughly 80 to 1; and why one-size-fits-all agent governance actually backfires.

Links

• Full transcript & framework: rickpollick.com/blog/valuestream-episode-3-the-agent-acts-you-answer
• Show home: rickpollick.com/podcast

Next episode
Boundaries got us here. Next time it's capability: giving the agent a brain, the context, memory, and priorities that turn a confidently-wrong assistant into something that actually knows your week.

Subscribe in your podcast app of choice. New episodes monthly. Companion essay on the blog with each ship.

This is Valuestream. I'm Rick Pollick. The agent acts, you answer.