What actually breaks
when you run LLM agents

Field notes from an operator who runs LLM agents on real work: tool permissions, silent failures, session handoffs, and the gates that catch them.

Spotted an error, or have a question?

Corrections and questions are welcome. Send them through the contact form.

Running agents and harnesses in practice: tool permissions, sub-agents that stop silently, pre-execution gates, and the incidents behind each rule.

How context is built and carried: an LLM-maintained wiki, session handoffs, ingest gates, and where progress state should and should not live.

Embeddings, retrieval, context windows. Covered only where a measurement or a failure from a real run makes the concept concrete.

No posts