Summary
Rémi Louf, CEO of .txt (a 15-person AI company), spent two weeks in early 2026 stepping away from day-to-day operations to build a fully autonomous morning briefing agent — one that would browse market news, review Linear/CRM tickets, and process long voice memos recorded during his morning walks, all without requiring him to babysit a terminal or phone. The talk at AI Engineer recounts that experience honestly, including the failures that shaped the architecture.
Louf’s central argument is that popular agent frameworks, while individually reasonable, tend to bury the most important part of an agent — the prompt — deep inside code, making iteration slow and version control awkward. His alternative: define agent behavior in plain YAML files that can be diffed, reviewed in PRs, and dropped into a folder for a lightweight runtime to pick up automatically. Scheduling via cron and triggered via standard tooling completes the primitives picture.
The failures he encountered map directly onto the runtime components he ended up building: a brief posted to Slack twice led to a proper task queue with attempt counting; a vanished voice note led to an append-only event log where everything is causally linked and queryable; a ruined market brief (caused by unversioned prompt changes) led to a content-addressed system inspired by git and Nix that makes every prompt change traceable. The session argues these are not novel inventions but well-understood distributed systems primitives — the insight is that agent runtimes need them just as much as any other production system does.
📺 Source: AI Engineer · Published August 22, 2026
🏷️ Format: Opinion Editorial
