--- schema: 1 type: video title: >- From 36% to 100%: How Self-Improving Agents Write Their Own Skills — Rafal Wilinski, Runlayer description: >- When one agent finally cracks a hard problem, what if every agent in the company learned it too? Rafal Wilinski, founding engineer at Runlayer, shows how to go from one self-improving agent to a self-improving company. He explains why skills matter more as agents work longer (a bad start can send them down an hours-long rabbit hole) and the three problems holding skills back: they're built for developers, nobody has time to write them, and many clients don't support them. His fix is to deliver skills over MCP from one governed server, then let agents write their own skills by distilling successful and failed runs, an idea borrowed from the Voyager Minecraft paper. In one test, a distilled skill took a task from 36% success to 100%. Multiply that across every agent and you get an organizational flywheel that keeps your hard-won knowledge even as models change. In this talk: • What skills are, and why they matter more as agents run longer • Using MCP as the distribution layer for skills across the whole company • Distilling skills from agent traces, and why failed runs teach the most • The organizational flywheel: knowledge competitors can't copy SPEAKER Rafal Wilinski, Founding Engineer, Runlayer LinkedIn: https://www.linkedin.com/in/rafwilinski X: https://x.com/rafalwilinski Website: https://rwilinski.ai LINKS Runlayer: https://www.runlayer.com Runlayer on X: https://x.com/runlayer CHAPTERS 0:00 Intro 0:32 Every hard problem becomes a playbook 1:37 About Runlayer 2:12 What a skill is 2:51 Agents can now work for hours 3:51 Longer tasks, deeper rabbit holes 4:31 Skills nobody wants to rediscover 5:21 Three problems with skills 6:01 The risk of installing random skills 7:25 Skills vs MCP 7:55 MCP as the distribution layer 8:40 The Skills Over MCP working group 10:00 Self-improving agents 10:10 Voyager's skill library 11:34 Distilling skills from traces 11:59 Failures teach the most 12:54 Result: 36% to 100% 14:04 Frontier intelligence is rented 15:23 The organizational flywheel 16:38 Your real moat Recorded at the AI Engineer World's Fair 2026 in San Francisco. Subscribe for more talks from the engineers building with AI. AI Engineer: https://ai.engineer YouTube: https://www.youtube.com/@aiDotEngineer X: https://x.com/aiDotEngineer LinkedIn: https://www.linkedin.com/company/aidotengineer/ #AIAgents #MCP #AIEngineer canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/from-36-to-100-how-self-improving-agents-write-thei-df0e0e40fb8d markdown: >- https://liquidgold.cc/channel/aidotengineer/video/from-36-to-100-how-self-improving-agents-write-thei-df0e0e40fb8d.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:19:42.184Z' language: null children: [] videoId: u-o0sW9nwmk durationSeconds: 1111 publishedAt: '2026-10-07T14:00:10.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 35 endSeconds: 45 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=35' heldReason: rights - ordinal: 1 text: null startSeconds: 145 endSeconds: 177 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=145' heldReason: rights - ordinal: 2 text: null startSeconds: 187 endSeconds: 233 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=187' heldReason: rights - ordinal: 3 text: null startSeconds: 246 endSeconds: 278 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=246' heldReason: rights - ordinal: 4 text: null startSeconds: 334 endSeconds: 367 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=334' heldReason: rights - ordinal: 5 text: null startSeconds: 378 endSeconds: 447 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=378' heldReason: rights - ordinal: 6 text: null startSeconds: 460 endSeconds: 482 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=460' heldReason: rights - ordinal: 7 text: null startSeconds: 493 endSeconds: 527 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=493' heldReason: rights - ordinal: 8 text: null startSeconds: 629 endSeconds: 696 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=629' heldReason: rights - ordinal: 9 text: null startSeconds: 731 endSeconds: 776 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=731' heldReason: rights - ordinal: 10 text: null startSeconds: 787 endSeconds: 853 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=787' heldReason: rights - ordinal: 11 text: null startSeconds: 853 endSeconds: 864 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=853' heldReason: rights - ordinal: 12 text: null startSeconds: 933 endSeconds: 944 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=933' heldReason: rights - ordinal: 13 text: null startSeconds: 999 endSeconds: 1010 watch: 'https://liquidgold.cc/watch/u-o0sW9nwmk?t=999' heldReason: rights --- # From 36% to 100%: How Self-Improving Agents Write Their Own Skills — Rafal Wilinski, Runlayer When one agent finally cracks a hard problem, what if every agent in the company learned it too? Rafal Wilinski, founding engineer at Runlayer, shows how to go from one self-improving agent to a self-improving company. He explains why skills matter more as agents work longer (a bad start can send them down an hours-long rabbit hole) and the three problems holding skills back: they're built for developers, nobody has time to write them, and many clients don't support them. His fix is to deliver skills over MCP from one governed server, then let agents write their own skills by distilling successful and failed runs, an idea borrowed from the Voyager Minecraft paper. In one test, a distilled skill took a task from 36% success to 100%. Multiply that across every agent and you get an organizational flywheel that keeps your hard-won knowledge even as models change. In this talk: • What skills are, and why they matter more as agents run longer • Using MCP as the distribution layer for skills across the whole company • Distilling skills from agent traces, and why failed runs teach the most • The organizational flywheel: knowledge competitors can't copy SPEAKER Rafal Wilinski, Founding Engineer, Runlayer LinkedIn: https://www.linkedin.com/in/rafwilinski X: https://x.com/rafalwilinski Website: https://rwilinski.ai LINKS Runlayer: https://www.runlayer.com Runlayer on X: https://x.com/runlayer CHAPTERS 0:00 Intro 0:32 Every hard problem becomes a playbook 1:37 About Runlayer 2:12 What a skill is 2:51 Agents can now work for hours 3:51 Longer tasks, deeper rabbit holes 4:31 Skills nobody wants to rediscover 5:21 Three problems with skills 6:01 The risk of installing random skills 7:25 Skills vs MCP 7:55 MCP as the distribution layer 8:40 The Skills Over MCP working group 10:00 Self-improving agents 10:10 Voyager's skill library 11:34 Distilling skills from traces 11:59 Failures teach the most 12:54 Result: 36% to 100% 14:04 Frontier intelligence is rented 15:23 The organizational flywheel 16:38 Your real moat Recorded at the AI Engineer World's Fair 2026 in San Francisco. Subscribe for more talks from the engineers building with AI. AI Engineer: https://ai.engineer YouTube: https://www.youtube.com/@aiDotEngineer X: https://x.com/aiDotEngineer LinkedIn: https://www.linkedin.com/company/aidotengineer/ #AIAgents #MCP #AIEngineer Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/from-36-to-100-how-self-improving-agents-write-thei-df0e0e40fb8d · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:12 to 1:40) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:40 to 2:15) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (2:15 to 2:57) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (2:57 to 3:53) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (3:53 to 5:24) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (5:24 to 7:27) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (7:27 to 10:04) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (10:04 to 11:36) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (11:36 to 14:13) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (14:13 to 18:31) > Held: transcript rights do not permit this chunk's text.