--- schema: 1 type: video title: >- Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind description: >- By the third version of the same agent, the source directory is gone. What remains is an AGENTS.md holding the system instruction and a small bash script that installs the GitHub CLI on first run. No tool definitions, no JSON schemas, no Python. Philipp Schmid builds the same GitHub pull request review agent three times on stage, deleting code with each pass, and the test he applies at the end of each one is the same throwaway question: what is the weather in San Francisco? Version one, a hand written loop with function call parsing and error handling, says it cannot do that. Version two, built on an agent framework that generates schemas from function signatures, says the same. Version three reaches for search and answers, because nobody told it what its tools were for. What makes the third version possible is a hosted sandbox the agent runs inside, with a network proxy that injects credentials on outbound requests so the agent never sees the token itself, and an allowed domain list you can narrow or leave open. Loops, tool routing, session state, and context compaction all move server side. What stays yours is the part Schmid argues was always the real work: instructions, rules, capabilities expressed as skill files, and the evaluations. His closing heuristic is worth the price of admission. If your harness gets more complex as models improve, you are probably overengineering it. He points at teams who have thrown out thousands of lines of orchestration in favor of a couple hundred lines of markdown, and at others rebuilding their harness several times a year. Speaker info: - https://x.com/_philschmid - https://www.linkedin.com/in/philipp-schmid-a6a2bb196/ - https://www.philschmid.de/ Timestamps: 0:00 - Three agents, less code each time 2:45 - Version one: the hand written Python loop 4:30 - Asking it something it has no tool for 5:21 - Version two: what a framework takes away 7:52 - Version three: a remote agent with a sandbox 8:42 - Credentials the agent never sees 12:09 - The weather question, answered at last 13:50 - What moves server side, what stays yours 16:22 - If your harness grows, you are overengineering 17:11 - Takeaways: stop micromanaging the model --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/fjF8EKnxKCU-agents-without-code-skills-yaml-filesystems 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/watch/fjF8EKnxKCU' markdown: 'https://liquidgold.cc/watch/fjF8EKnxKCU.md' llms: 'https://liquidgold.cc/watch/fjF8EKnxKCU/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T03:40:42.432Z' language: null children: [] videoId: fjF8EKnxKCU durationSeconds: 1108 publishedAt: '2026-09-14T17:30:25.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 25 endSeconds: 169 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=25' heldReason: rights - ordinal: 1 text: null startSeconds: 25 endSeconds: 169 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=25' heldReason: rights - ordinal: 2 text: null startSeconds: 181 endSeconds: 279 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=181' heldReason: rights - ordinal: 3 text: null startSeconds: 321 endSeconds: 334 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=321' heldReason: rights - ordinal: 4 text: null startSeconds: 334 endSeconds: 472 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=334' heldReason: rights - ordinal: 5 text: null startSeconds: 482 endSeconds: 526 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=482' heldReason: rights - ordinal: 6 text: null startSeconds: 526 endSeconds: 536 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=526' heldReason: rights - ordinal: 7 text: null startSeconds: 482 endSeconds: 526 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=482' heldReason: rights - ordinal: 8 text: null startSeconds: 536 endSeconds: 729 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=536' heldReason: rights - ordinal: 9 text: null startSeconds: 536 endSeconds: 729 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=536' heldReason: rights - ordinal: 10 text: null startSeconds: 729 endSeconds: 741 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=729' heldReason: rights - ordinal: 11 text: null startSeconds: 741 endSeconds: 834 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=741' heldReason: rights - ordinal: 12 text: null startSeconds: 844 endSeconds: 986 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=844' heldReason: rights - ordinal: 13 text: null startSeconds: 844 endSeconds: 986 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=844' heldReason: rights - ordinal: 14 text: null startSeconds: 986 endSeconds: 997 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=986' heldReason: rights - ordinal: 15 text: null startSeconds: 997 endSeconds: 1031 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=997' heldReason: rights - ordinal: 16 text: null startSeconds: 1031 endSeconds: 1041 watch: 'https://liquidgold.cc/watch/fjF8EKnxKCU?t=1031' heldReason: rights --- # Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind By the third version of the same agent, the source directory is gone. What remains is an AGENTS.md holding the system instruction and a small bash script that installs the GitHub CLI on first run. No tool definitions, no JSON schemas, no Python. Philipp Schmid builds the same GitHub pull request review agent three times on stage, deleting code with each pass, and the test he applies at the end of each one is the same throwaway question: what is the weather in San Francisco? Version one, a hand written loop with function call parsing and error handling, says it cannot do that. Version two, built on an agent framework that generates schemas from function signatures, says the same. Version three reaches for search and answers, because nobody told it what its tools were for. What makes the third version possible is a hosted sandbox the agent runs inside, with a network proxy that injects credentials on outbound requests so the agent never sees the token itself, and an allowed domain list you can narrow or leave open. Loops, tool routing, session state, and context compaction all move server side. What stays yours is the part Schmid argues was always the real work: instructions, rules, capabilities expressed as skill files, and the evaluations. His closing heuristic is worth the price of admission. If your harness gets more complex as models improve, you are probably overengineering it. He points at teams who have thrown out thousands of lines of orchestration in favor of a couple hundred lines of markdown, and at others rebuilding their harness several times a year. Speaker info: - https://x.com/\_philschmid - https://www.linkedin.com/in/philipp-schmid-a6a2bb196/ - https://www.philschmid.de/ Timestamps: 0:00 - Three agents, less code each time 2:45 - Version one: the hand written Python loop 4:30 - Asking it something it has no tool for 5:21 - Version two: what a framework takes away 7:52 - Version three: a remote agent with a sandbox 8:42 - Credentials the agent never sees 12:09 - The weather question, answered at last 13:50 - What moves server side, what stays yours 16:22 - If your harness grows, you are overengineering 17:11 - Takeaways: stop micromanaging the model --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/fjF8EKnxKCU-agents-without-code-skills-yaml-filesystems 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/watch/fjF8EKnxKCU · 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.