--- schema: 1 type: video title: 'Stop Rationing Tokens: Let the Harness Pick the Model — Kimchi by Cast AI' description: >- Token budgets are out of control, and limiting tokens cripples your developers. Cast AI took a different route. Laurent Gil, co-founder and president of Cast AI, and Žilvinas Urbonas, who leads engineering for Kimchi, show how Cast AI built its own open-source coding harness after its coding-agent bill took off. The key idea is to measure cost per task instead of cost per token, and to let the harness choose the right proprietary or open model for each task based on the outcome. They share three months of internal results with 300 employees, then demo Ferment for multi-hour autonomous tasks, Teleport for agent sessions that keep running after you close your laptop, and Studio for team collaboration. In this talk: • Why cost per task matters more than cost per token • How an outcome-aware harness switches models automatically as new ones ship • Ferment: milestone-based, self-scoring autonomous coding runs that deploy to staging • Teleport and Studio: remote sandboxes and a shared Kanban board for agent sessions SPEAKERS Laurent Gil, Co-founder & President, Cast AI LinkedIn: https://www.linkedin.com/in/laurentgil/ X: https://x.com/laurentgil Žilvinas Urbonas, Engineering lead, Kimchi (Cast AI) LinkedIn: https://www.linkedin.com/in/zilvinasurbonas/ LINKS Kimchi: https://kimchi.dev/ Kimchi on X: https://x.com/getkimchi Kimchi (GitHub): https://github.com/getkimchi/kimchi Cast AI: https://cast.ai CHAPTERS 0:00 Intro 0:13 Meet Kimchi 0:52 Token costs out of control 1:31 Why rationing tokens is the wrong answer 2:26 Make tokens unlimited and cheap 3:05 Cost per token vs cost per task 4:30 A harness that picks the model 4:49 Results: 2.5x savings in 3 months 5:54 Watching the harness switch models 7:23 How the harness works 8:15 Ferment: multi-hour autonomous tasks 9:05 Build, verify, deploy to staging 11:03 Reading diffs isn't enough anymore 11:48 Kimchi Teleport 14:02 Why 62% of engineers code this way 14:23 Kimchi Studio for teams 17:04 Wrap-up 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/ #CodingAgents #OpenSourceAI #AIEngineer canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/stop-rationing-tokens-let-the-harness-pick-the-mode-5af5aca9943b markdown: >- https://liquidgold.cc/channel/aidotengineer/video/stop-rationing-tokens-let-the-harness-pick-the-mode-5af5aca9943b.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:20:10.380Z' language: null children: [] videoId: 48YUYDjwfYY durationSeconds: 1088 publishedAt: '2026-10-02T20:00:14.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 59 endSeconds: 69 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=59' heldReason: rights - ordinal: 1 text: null startSeconds: 69 endSeconds: 102 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=69' heldReason: rights - ordinal: 2 text: null startSeconds: 102 endSeconds: 113 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=102' heldReason: rights - ordinal: 3 text: null startSeconds: 113 endSeconds: 148 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=113' heldReason: rights - ordinal: 4 text: null startSeconds: 148 endSeconds: 158 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=148' heldReason: rights - ordinal: 5 text: null startSeconds: 158 endSeconds: 191 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=158' heldReason: rights - ordinal: 6 text: null startSeconds: 191 endSeconds: 201 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=191' heldReason: rights - ordinal: 7 text: null startSeconds: 201 endSeconds: 270 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=201' heldReason: rights - ordinal: 8 text: null startSeconds: 270 endSeconds: 281 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=270' heldReason: rights - ordinal: 9 text: null startSeconds: 281 endSeconds: 292 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=281' heldReason: rights - ordinal: 10 text: null startSeconds: 292 endSeconds: 302 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=292' heldReason: rights - ordinal: 11 text: null startSeconds: 302 endSeconds: 359 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=302' heldReason: rights - ordinal: 12 text: null startSeconds: 359 endSeconds: 370 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=359' heldReason: rights - ordinal: 13 text: null startSeconds: 370 endSeconds: 453 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=370' heldReason: rights - ordinal: 14 text: null startSeconds: 453 endSeconds: 464 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=453' heldReason: rights - ordinal: 15 text: null startSeconds: 464 endSeconds: 497 watch: 'https://liquidgold.cc/watch/48YUYDjwfYY?t=464' heldReason: rights --- # Stop Rationing Tokens: Let the Harness Pick the Model — Kimchi by Cast AI Token budgets are out of control, and limiting tokens cripples your developers. Cast AI took a different route. Laurent Gil, co-founder and president of Cast AI, and Žilvinas Urbonas, who leads engineering for Kimchi, show how Cast AI built its own open-source coding harness after its coding-agent bill took off. The key idea is to measure cost per task instead of cost per token, and to let the harness choose the right proprietary or open model for each task based on the outcome. They share three months of internal results with 300 employees, then demo Ferment for multi-hour autonomous tasks, Teleport for agent sessions that keep running after you close your laptop, and Studio for team collaboration. In this talk: • Why cost per task matters more than cost per token • How an outcome-aware harness switches models automatically as new ones ship • Ferment: milestone-based, self-scoring autonomous coding runs that deploy to staging • Teleport and Studio: remote sandboxes and a shared Kanban board for agent sessions SPEAKERS Laurent Gil, Co-founder & President, Cast AI LinkedIn: https://www.linkedin.com/in/laurentgil/ X: https://x.com/laurentgil Žilvinas Urbonas, Engineering lead, Kimchi (Cast AI) LinkedIn: https://www.linkedin.com/in/zilvinasurbonas/ LINKS Kimchi: https://kimchi.dev/ Kimchi on X: https://x.com/getkimchi Kimchi (GitHub): https://github.com/getkimchi/kimchi Cast AI: https://cast.ai CHAPTERS 0:00 Intro 0:13 Meet Kimchi 0:52 Token costs out of control 1:31 Why rationing tokens is the wrong answer 2:26 Make tokens unlimited and cheap 3:05 Cost per token vs cost per task 4:30 A harness that picks the model 4:49 Results: 2.5x savings in 3 months 5:54 Watching the harness switch models 7:23 How the harness works 8:15 Ferment: multi-hour autonomous tasks 9:05 Build, verify, deploy to staging 11:03 Reading diffs isn't enough anymore 11:48 Kimchi Teleport 14:02 Why 62% of engineers code this way 14:23 Kimchi Studio for teams 17:04 Wrap-up 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/ #CodingAgents #OpenSourceAI #AIEngineer Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/stop-rationing-tokens-let-the-harness-pick-the-mode-5af5aca9943b · 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.