--- schema: 1 type: video title: 'Why We Deleted Our MCP Server and Rebuilt It — Abhi Arya, Reducto' description: >- Reducto gave its whole team an MCP server with a tool for every endpoint. It broke the same day. Abhi Arya, who works on product at Reducto, argues that building agent-first software comes down to one question: who pays when the agent is wrong? He describes three approaches, auto mode, user-first, and agent experience in service of the user, and how Reducto learned the difference. A first MCP server with one tool per API endpoint demoed well but fell apart when the whole team used it: the agent built entire workflows confidently, and nobody could explain them. He deleted it in one commit and rebuilt with fewer tools that carry the structure of real work, plus a snapshot tool for live state. Then he made uncertainty visible, with confidence scores and an agent that asks instead of guessing, and added instrumentation that made the system improve on its own, until the sales team adopted it without being asked. In this talk: • Auto mode vs user-first vs agent experience in service of the user • Why one tool per API endpoint fails, and what to build instead • Making uncertainty visible: confidence scores and asking instead of guessing • Instrumentation that turns usage into a self-improving loop SPEAKER Abhi Arya, Product, Reducto LinkedIn: https://www.linkedin.com/in/abhiaarya/ X: https://x.com/abhiarya GitHub: https://github.com/abhi-arya1 Website: https://abhi-arya1.github.io LINKS Reducto: https://reducto.ai Reducto MCP server docs: https://docs.reducto.ai/mcp-server Reducto on X: https://x.com/reductoai CHAPTERS 0:00 Intro 0:38 3 billion documents 1:48 What is agent-first software? 2:38 Who pays when the agent is wrong? 2:48 Auto mode, user-first, or in service of the user 3:57 Stuck in configuration 4:47 Version 1: a tool for every endpoint 5:27 It broke immediately 6:27 Fewer tools that carry real structure 7:32 A snapshot of live state 8:17 Don't just hand agents an API 9:07 Failures that happen quietly 9:42 Make uncertainty visible 10:22 Ask, don't guess 11:42 One root cause 13:07 Instrumentation and a self-improving loop 14:12 Sales adopted it on their own 15:12 What agent-first really means 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/ #AgentExperience #MCP #AIEngineer canonical: 'https://liquidgold.cc/watch/jJQoVkd5yLg' markdown: 'https://liquidgold.cc/watch/jJQoVkd5yLg.md' llms: 'https://liquidgold.cc/watch/jJQoVkd5yLg/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:19:33.327Z' language: null children: [] videoId: jJQoVkd5yLg durationSeconds: 1000 publishedAt: '2026-10-08T17:00:23.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 124 endSeconds: 158 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=124' heldReason: rights - ordinal: 1 text: null startSeconds: 158 endSeconds: 169 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=158' heldReason: rights - ordinal: 2 text: null startSeconds: 179 endSeconds: 246 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=179' heldReason: rights - ordinal: 3 text: null startSeconds: 257 endSeconds: 293 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=257' heldReason: rights - ordinal: 4 text: null startSeconds: 349 endSeconds: 394 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=349' heldReason: rights - ordinal: 5 text: null startSeconds: 404 endSeconds: 456 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=404' heldReason: rights - ordinal: 6 text: null startSeconds: 456 endSeconds: 468 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=456' heldReason: rights - ordinal: 7 text: null startSeconds: 501 endSeconds: 513 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=501' heldReason: rights - ordinal: 8 text: null startSeconds: 559 endSeconds: 584 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=559' heldReason: rights - ordinal: 9 text: null startSeconds: 584 endSeconds: 596 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=584' heldReason: rights - ordinal: 10 text: null startSeconds: 629 endSeconds: 641 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=629' heldReason: rights - ordinal: 11 text: null startSeconds: 712 endSeconds: 722 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=712' heldReason: rights - ordinal: 12 text: null startSeconds: 722 endSeconds: 790 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=722' heldReason: rights - ordinal: 13 text: null startSeconds: 800 endSeconds: 852 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=800' heldReason: rights - ordinal: 14 text: null startSeconds: 862 endSeconds: 916 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=862' heldReason: rights - ordinal: 15 text: null startSeconds: 927 endSeconds: 1000 watch: 'https://liquidgold.cc/watch/jJQoVkd5yLg?t=927' heldReason: rights --- # Why We Deleted Our MCP Server and Rebuilt It — Abhi Arya, Reducto Reducto gave its whole team an MCP server with a tool for every endpoint. It broke the same day. Abhi Arya, who works on product at Reducto, argues that building agent-first software comes down to one question: who pays when the agent is wrong? He describes three approaches, auto mode, user-first, and agent experience in service of the user, and how Reducto learned the difference. A first MCP server with one tool per API endpoint demoed well but fell apart when the whole team used it: the agent built entire workflows confidently, and nobody could explain them. He deleted it in one commit and rebuilt with fewer tools that carry the structure of real work, plus a snapshot tool for live state. Then he made uncertainty visible, with confidence scores and an agent that asks instead of guessing, and added instrumentation that made the system improve on its own, until the sales team adopted it without being asked. In this talk: • Auto mode vs user-first vs agent experience in service of the user • Why one tool per API endpoint fails, and what to build instead • Making uncertainty visible: confidence scores and asking instead of guessing • Instrumentation that turns usage into a self-improving loop SPEAKER Abhi Arya, Product, Reducto LinkedIn: https://www.linkedin.com/in/abhiaarya/ X: https://x.com/abhiarya GitHub: https://github.com/abhi-arya1 Website: https://abhi-arya1.github.io LINKS Reducto: https://reducto.ai Reducto MCP server docs: https://docs.reducto.ai/mcp-server Reducto on X: https://x.com/reductoai CHAPTERS 0:00 Intro 0:38 3 billion documents 1:48 What is agent-first software? 2:38 Who pays when the agent is wrong? 2:48 Auto mode, user-first, or in service of the user 3:57 Stuck in configuration 4:47 Version 1: a tool for every endpoint 5:27 It broke immediately 6:27 Fewer tools that carry real structure 7:32 A snapshot of live state 8:17 Don't just hand agents an API 9:07 Failures that happen quietly 9:42 Make uncertainty visible 10:22 Ask, don't guess 11:42 One root cause 13:07 Instrumentation and a self-improving loop 14:12 Sales adopted it on their own 15:12 What agent-first really means 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/ #AgentExperience #MCP #AIEngineer Canonical: https://liquidgold.cc/watch/jJQoVkd5yLg · 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 1:54) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:54 to 2:38) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (2:38 to 4:06) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (4:06 to 5:03) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (5:03 to 6:34) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (6:34 to 8:21) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (8:21 to 9:19) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (9:19 to 10:29) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (10:29 to 11:52) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (11:52 to 13:10) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (13:10 to 14:12) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (14:12 to 15:16) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (15:16 to 16:40) > Held: transcript rights do not permit this chunk's text.