
Roll Your Own Agent Harness: Building Daiku and Personal Agents with Tejas Kumar "Let me send a prompt here, 'Delete everything.' That's what I did on stage. And it actually just did it." @TejasKumar_, AI Engineer and author of Fluent React, talked to us at @aiDotEngineer CODE in SF about his AIE CODE workshop, building agent harnesses from scratch, and why "can the agent call this tool" and "should it call this tool right now" are two different questions. We get into: > Why banks, insurers, and hospitals are building their own harnesses instead of using off-the-shelf coding agents > The guardrails that came after rm -rf: allow lists, ask-first tools, and Face ID before sensitive actions > Getting the right human in the loop, not the whole team > Context caching, the piece most self-built harnesses are missing > Daiku, a personal agent harness that isn't locked to one model vendor TIMESTAMPS (00:40) The workshop: from an empty while loop to a coding agent harness (01:45) Rolling your own harness for air-gapped enterprise infrastructure (02:40) Why regulated industries can't just use off-the-shelf harnesses (03:40) Can you call the tool vs should you call it right now (04:10) "Delete everything" on stage, and the guardrails that came next (04:50) The right human in the loop (05:50) Context caching and why self-built harnesses overpay for tokens (07:20) The Resy incident: an agent that tried too hard (07:55) AAuth and budgets for downstream resources (09:50) Keeping credentials out of the agent and the deployment (11:15) The power plug you can pull on a rogue agent (12:05) Daiku: one computer every one of your agents can drive (13:05) The agent that signed Tejas up for Shark Tank (15:45) Switching models mid task when you run out of usage













