project
A blog publishing field notes and technical essays on designing, debugging, and operating LLM-based agents reliably in production.
No evidence-backed relationships are recorded.
A field-notes post on production LLM agents argues the context window behaves like a cache under eviction pressure, and that the real work is deciding every turn what earns space and which failed turns to scrub.
Reality
A field-notes post argues the metric that matters is how often an agent fails on inputs you did not choose, and that most of that rate is forecastable in a harness before release.
Reality
A dev.to post sets out three fixed pipeline shapes for tasks that get an agent loop by default. Its real test is whether every possible run can be drawn in advance.
Reality
A dev.to field-notes post lays out how partial_json actually arrives, and why the three common ways of handling it fail in three different places.
Reality
A field-notes post argues the retry question splits in two: how many, and whether at all. On a user-facing path with a 30-second ceiling, the honest answer is usually a fast error.
Reality
A Loop & Retry field note argues idempotency keys and per-attempt traces are preconditions for a retry budget. Its own sample key construction shows how easily that gets wrong.
Reality
A dev.to implementation note ports one token-bucket retry budget into Python, Go, and JavaScript. The Python trap is a missing lock; the Go trap is the select.
Reality