--- schema: 1 type: video title: 'Agent Frameworks Considered Harmful — Rémi Louf, .txt' 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 - https://www.linkedin.com/in/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 --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/KHudyx5wW3U-youtube-62bf13b3b1ccf95de497 Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/agent-frameworks-considered-harmful-remi-louf-txt-e50339962c83 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/agent-frameworks-considered-harmful-remi-louf-txt-e50339962c83.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T04:50:45.155Z' language: null children: [] videoId: KHudyx5wW3U durationSeconds: 1229 publishedAt: '2026-08-22T16:30:39.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 22 endSeconds: 124 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=22' heldReason: rights - ordinal: 1 text: null startSeconds: 134 endSeconds: 169 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=134' heldReason: rights - ordinal: 2 text: null startSeconds: 181 endSeconds: 272 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=181' heldReason: rights - ordinal: 3 text: null startSeconds: 283 endSeconds: 372 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=283' heldReason: rights - ordinal: 4 text: null startSeconds: 372 endSeconds: 382 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=372' heldReason: rights - ordinal: 5 text: null startSeconds: 382 endSeconds: 533 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=382' heldReason: rights - ordinal: 6 text: null startSeconds: 382 endSeconds: 533 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=382' heldReason: rights - ordinal: 7 text: null startSeconds: 578 endSeconds: 589 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=578' heldReason: rights - ordinal: 8 text: null startSeconds: 589 endSeconds: 733 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=589' heldReason: rights - ordinal: 9 text: null startSeconds: 733 endSeconds: 745 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=733' heldReason: rights - ordinal: 10 text: null startSeconds: 745 endSeconds: 935 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=745' heldReason: rights - ordinal: 11 text: null startSeconds: 935 endSeconds: 946 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=935' heldReason: rights - ordinal: 12 text: null startSeconds: 989 endSeconds: 999 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=989' heldReason: rights - ordinal: 13 text: null startSeconds: 999 endSeconds: 1089 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=999' heldReason: rights - ordinal: 14 text: null startSeconds: 1099 endSeconds: 1229 watch: 'https://liquidgold.cc/watch/KHudyx5wW3U?t=1099' heldReason: rights --- # Agent Frameworks Considered Harmful — Rémi Louf, .txt 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 - https://www.linkedin.com/in/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 --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/KHudyx5wW3U-youtube-62bf13b3b1ccf95de497 Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/agent-frameworks-considered-harmful-remi-louf-txt-e50339962c83 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.