Field notes

Notes from the control tower.

Short essays on agent operations, harnesses, evidence, and what we are learning while building ATC in public.

the-harness-is-the-product

The harness is the product

Models matter. The operating environment decides whether agent work is accountable.

Agent teams do not fail only because the model got confused. They fail because the work has no durable operating surface: no assignment, no evidence gate, no recovery path, no shared record of what shipped.

ATC is the control plane for that surface. Dispatch turns intent into lanes. Concourse shows work in motion. Hangar manages the workforce. Flight Logs preserve the black-box record. The harness becomes visible to the operator.

plan-work-evidence-outcome

Plan -> Work -> Evidence -> Outcome

A simple loop for making autonomous work legible.

ATC keeps the loop short on purpose. A Flight Plan creates lanes and waypoints. Workers claim work, attach evidence, and complete tasks. Operators merge lanes only when evidence exists. Outcomes answer whether shipped work mattered.

This loop is more important than a chat transcript. The transcript can stay in the worker. ATC stores the canonical record operators need to trust the work.

what-atc-is-not

What ATC is not

Not a chat box. Not a wrapper. Not one more dashboard.

ATC is not trying to be the agent. It is the tower. Agents can run in OpenClaw, Cursor SDK, Claude Code, Codex, or something that has not shipped yet.

The point is that operators get one place to see assignment, state, evidence, recovery, and outcome across those workers.