AI agents: From pair programming to AFK programming
Field notes on getting coding agents to build, validate, and ship, without you watching every step.
If you run Claude Code, Codex, or OpenCode, you already know the limit. The agents got faster, but the loop is still bottlenecked by you. You're still saying "do this, now do that, run that, try again," and you're still watching every step.
Anil Dukkipatty builds Grass (cloud infra for autonomous coding agents) and ships its features end to end this way: the agent picks up a vertical slice, builds it, validates it against acceptance criteria, fixes its own failures, and reports back with evidence. His job shifted from writing and supervising to defining the loop. He calls it AFK programming.
"Don't prompt the agent to code. Set the agent up to operate."
When the agent runs without you watching, the trace it leaves is the only record of what actually happened. Tarun Sachdeva, founder of Traces, joins for the second half on why those traces are worth keeping: how a saved trace turns one agent's run into something your whole team can read, replay, and build on, what he calls multiplayer AI. He'll demo Traces live on the sessions from this workshop.
Who this is for
Engineers running Claude Code, Codex, or OpenCode daily who are tired of babysitting
Founders shipping with one or more agents in parallel
Anyone who's tried "set it and forget it" and watched the agent walk off a cliff
What you'll walk away with
An alignment-spec workflow that catches misunderstanding before any code is written
How to break work into PR-sized vertical slices an agent can actually finish
The builder-validator loop pattern, with a concrete example you can copy
A reusable-skills setup so you stop re-explaining your workflow every run
A working mental model for what to delegate and what to keep
How to store and share agent traces so a run stays reviewable instead of getting lost in a terminal
Why shared traces turn solo agent runs into multiplayer AI your team can build on
Format
75 minutes. Anil walks the AFK loop on a real project, then Tarun walks through storing and sharing traces with a live Traces demo, then shared Q&A. Drop a task you want to delegate in the registration form and Anil will pick one to wire up live. Recorded, so signups get the recording either way.
What you'll need
Familiarity with at least one of Claude Code, Codex, or OpenCode
A real task you want to delegate (small but complete is fine)
Attendees also get beta access to Grass, the infra Anil uses to run these loops, with sessions pre-configured for the workflow above.
About Anil
Anil Dukkipatty is co-founder and CTO of Revise, the team behind Grass. He architected Grass's core infra: real-time sync between agent VMs and the iOS app, SSE event buffering, and Daytona VM orchestration. He now ships Grass features end to end using the AFK loop he'll walk through.
About Tarun
Tarun Sachdeva is founder and CEO of Traces, a platform for sharing and collaborating on coding-agent traces. He builds the tooling that lets teams see what their agents did and pick up where another run left off.