A field guide to building reliable AI agents

The Agent Loop is a curated knowledge base and Q&A site for engineers building AI agents. Every entry is a sourced field note — a durable pattern, a measured result, or a lesson from a public post-mortem — kept short enough to read in a minute and cited so you can verify it.

The name is the operating model: read → act → observe → correct. Reliable agents are loops with clear goals, useful tools, bounded context, and feedback — the model is only one component.

The site is run by the team behind The Agent Loop YouTube channel. The knowledge base itself is open (CC BY 4.0) and maintained in the agent-loop-kb repository — every entry carries a stable id, a canonical source URL, an explicit grade, and a last-verified date, all enforced by CI. Grading is honest by design: entries start unrated and earn an evidence-backed A–D grade as they are reviewed; the grade and its evidence ship in the public feed, so you can see exactly which entries have been verified and which are still pending review.

How entries are curated

• Every entry links a canonical source — research posts, official docs, papers, and public post-mortems — so you can verify it yourself.
• Each entry states what is known, where it came from, and when it was last verified. Long-form deep dives cite their primary sources inline.
• Entries older than 90 days without re-verification are flagged stale by CI.
• Corrections and contributions are welcome via issues or pull requests on the KB repository.

Contact

Found an error, want to suggest an entry, or need to reach us? Open an issue on the KB repository or leave a comment on any episode of the YouTube channel — both are checked regularly.

Looking for the answers themselves? Browse the library or ask the loop.