Steering Agents with Policies: an engineering session

Most agent failures aren't the model being bad at the task. The agent is missing one fact at one moment: the server dies when the shell closes, the file it's about to overwrite is the only copy, the output is right except for one byte. A policy that notices that moment and tells the agent what it's missing changes the outcome, without touching the model.

Nikita Agarwal (https://x.com/Niki_agi, co-founder and CEO of Failproof AI and co-author of the FIRE paper, will walk through how we write policies that steer agents at runtime, and what we learned from reading thousands of failed runs.

This is an engineering session, not a product demo. Here's what you'll get into

From failed run to policy. How we read failed traces, find the mistake that keeps coming back, and turn it into a policy. With real examples from coding agents.

Allow, deny, instruct. When to block a tool call, when to let it through, and when to send the agent an instruction instead. Why instruct ends up doing most of the work.

Why the wording matters. In our experiments, a specific instruction at the right moment beat a generic "verify your work" at the same moment. Nikita will show what a good instruction looks like and what a bad one looks like.

When policies go wrong. Policies that fire too often, policies that never match reworded tasks, and a policy set that made runs almost 50% more expensive. What we changed.

Writing your first one. A short live example of writing and testing a policy against a real failure, so you can do the same on your own agents.

Your questions, live. The last 20 minutes are open. Bring the failure your agents keep hitting.

Schedule

• 6:00 · why agents fail and where policies fit
• 6:15 · from traces to policies, with real examples
• 6:40 · open Q&A

For engineers running agents in production or building agent harnesses, who want a practical way to make them more reliable without retraining.

Tuesday, October 27 · 6:00 to 7:00 pm