← back
Agent Frameworks Considered Harmful — Rémi Louf, .txt
Takeaway
Simple event-driven agent definitions still require a reliable runtime that preserves history and handles distributed-system failures.
Summary
- Rémi Louf builds an unattended personal assistant for market updates, voice-note processing, and a daily briefing using declarative agent files.
- Cron handles scheduled work, while event subscriptions trigger agents when voice notes, emails, CRM changes, or pull requests arrive.
- Agents declare accepted and emitted events, allowing workflows to compose without manually maintaining orchestration graph edges.
- Lost notes, duplicate briefings, and untracked prompt changes expose the need for durable logs, reliable retry handling, and versioned configuration.
agent-runtimeevent-drivenorchestration
Original description
In the first week the daily brief posted to Slack twice, a voice note vanished entirely, and the market brief turned to garbage after prompt edits Rémi Louf had not versioned and could no longer recall. Each failure became a piece of what turned into a runtime. The lost note became an append only log where nothing is discarded and every event is causally linked to the one that triggered it. The duplicates became a real queue that counts attempts. The untraceable prompt became a content addressed store, and that one was the rabbit hole. Every part of a prompt is hashed and stored separately, the system message, each skill description, each tool definition, the user question, so a prompt is a list of hashes rather than a rendered string. Diff two runs and you see exactly which component changed. Replay one against a different model and the request rebuilds from the graph. He wanted this because a live chat session does not show you what the model actually saw, not with compaction and unshared reasoning in play. The rest stays deliberately small. Agents are markdown files you drop in a folder, so people who do not write code can add one. They subscribe to events instead of living in a graph with edges to maintain. Typed tool calls and typed events are the two boundaries, because roughly 20% of his events were coming back malformed and getting rejected. He took two weeks away from running a 15 person company to build it. Twenty agents now run there. Speaker info: https://x.com/remilouf / remilouf https://thetypicalset.com Timestamps: 0:00 - Two weeks away to find out what agents can do 1:57 - The morning he wanted, and the robot mower 2:49 - A terminal you still have to sit at 4:30 - Why the prompt ended up in markdown, not code 6:11 - Cron gets you when, events get you because 8:44 - What broke in week one 9:37 - The log: nothing lost, everything causally linked 12:11 - Content addressing the prompt, and free diffs and replays 15:32 - A kernel, not a framework 16:25 - Typed calls and typed events as boundaries 18:05 - Lessons, and build before you buy