Field Notes · No. 01
The cost of a file nobody reads
Field Note 01 opens Agent Centered Design — Build TBD's research programme translating Lean's discipline of flow and waste to agent operations. The first waste: over-production.
29 August 2026 · 4 min read
The programme
Build TBD is opening its research
The programme is called Agent Centered Design: Lean manufacturing's treatment of flow and waste, translated to agent operations. One rule governs it — every claim is measured in a live environment, or it does not ship.
Field Note 01 is about over-production, the waste Taiichi Ohno ranked worst because it conceals the others.
The environment
What follows is not a system running badly
The environment sets the bar. Command Center is our agentic delivery system: work goes to agents as briefed packages — scope, runbook, prior patterns assembled up front — rather than being negotiated turn by turn. 86.7% of dispatched work completes without a single human turn. Per unit delivered, it takes 11× less human interchange and about 3× fewer tokens than the same work done conversationally (a dispatched turn carried 51,700 tokens of context on average; a conversational turn carried 188,000).
So what follows is not the waste of a system running badly. It is what remains after a system runs well.
measured, our operation
The ledger
What the agents consumed
What the agents consumed
839 sessions · 26,274 API requests · 12 codebases · 2026-07-01 → 2026-08-29
| 176M | tokens written into context |
|---|---|
| 6,646M | re-read from cache on later turns |
| 15M | generated by the models |
| 6,837M | total — 6.84 billion tokens |
The window holds 1,130 of 2,163 recorded prompts; every figure is a floor.
measured, our operation
The finding
Authored, and never touched again
The records the agents authored are the deliberate pile: plans, logs, retrospectives, handover manifests, telemetry. No build artifacts, no dependency trees. And four internal stores — a swarm log, a retrospective record, a dispatch manifest, tool telemetry — had zero readers between them.
Records the agents authored
2,460
377
records authored
never touched again
15% never touched again, on the generous count — no read, no grep, no command.
measured, our operation
The loop
One box changes. Nothing else does.
An agent finishes a piece of work and writes a record of it. The record lands in a store. Nothing reads the store, so the question the record was written to answer is never answered — and the next record lands on top of it. Here is that loop, and here is the same loop with one box added.
The fix
We did not write less. We built readers.
The fix was not writing less. We built readers for those four stores. Records nothing needed fell from 17.8% to 12.2% — a third fewer — without touching the write path.
measured, our operation
In a factory you can see over-production stacked against the wall. In an agent system it is a well-written file in a well-named folder, authored by something doing excellent work.
Why the name
Reading has to be designed
Writing is nearly free for an agent; reading has to be designed. That asymmetry is why the programme is called Agent Centered Design — the system, not the agent, determines what gets read. Field notes will publish as they are measured.
If you run agents, the diagnostic costs five minutes: pick a store your system writes to, and ask what reads it.