Story · Rémi Louf (.txt)
Agent Frameworks Considered Harmful (Rémi Louf (.txt))
talk · Story page
A small runtime built from production failures: an append-only causal event log, a queue that counts attempts, and a content-addressed store that hashes every prompt component separately. Diff two runs to see exactly which component changed, or replay one against a different model.
In plain words
- Rémi Louf built a small system to organize work done by action-taking artificial intelligence after duplicate posts, lost notes, and unexplained quality drops.
- An unchangeable event record preserves what happened and links each action to the event that caused it.
- A work queue counts attempts, helping the system track retries instead of accidentally repeating or losing jobs.
- Each instruction part receives a unique code based on its contents, letting people compare runs or rebuild the same request elsewhere.
- Teams can trace failures and repeat earlier runs even when live conversations hide or shorten what the system previously received.
Appeared in
- Poisoned agent memory beats screening, and prompt rules keep failing as boundaries
Aug 25, 2026 · in the sections
Subscribe
Get the brief in your inbox
Pick daily, weekly, or both. Nothing is gated either way: every issue is on the site and in the feeds.
- Weekdays at 8:45am IST, one lead story and 6 to 9 items.
- Sundays, an argued synthesis rather than a recap.
- One click to leave, and quiet days say so in the subject line.