Reading
Recommended Reading
Things I found worth the time and expect to point at again. Every title links out to the source — one sentence here on why it's worth yours.
- Unrolling the Codex agent loop
Michael Bolin · Feb 2026
A first-party walk through the loop itself — prompt, inference, tool calls, and the assistant message that ends a turn — rather than another account of what a coding agent can do.
- Effective harnesses for long-running agents
Justin Young · Nov 2025
Treats "the agent forgets everything when the context window ends" as a harness problem rather than a model problem, and names the two failure modes it produces — doing too much at once, and declaring done too early.