Skip to content
tensorhaze

Menu

Motion

Reduced motion shows still images instead of animation. Everything else stays the same.

Study 04 · Working method

Agent Fleets

How one person runs AI coding agents on real projects. Every claim needs evidence, and the reviewer never sees the plan.

  • Method in use since Sep 2026
The loop: brief, spec, slices, evidence, reviewA brief becomes a spec and plan, then small slices of work, then an evidence folder, then a blind review. The reviewer receives the brief and the result. The line from the spec and plan to the reviewer stops short: the reviewer never sees them.1BriefR1 to Rn2Spec and planTraced to brief3SlicesOne change each4EvidenceProof per claim5Blind reviewBrief and resultSpec and plan stay hidden The loop: brief, spec, slices, evidence, reviewA brief becomes a spec and plan, then small slices of work, then an evidence folder, then a blind review. The reviewer receives the brief and the result. The line from the spec and plan to the reviewer stops short: the reviewer never sees them.1BriefR1 to Rn2Spec and planTraced to brief3SlicesOne change each4EvidenceProof per claim5Blind reviewBrief and result

01

The loop

Five steps, each leaving something a reviewer can check.

  1. BriefThe person’s request is split word for word into numbered requirements, R1 to Rn.1
  2. Spec and planEach requirement is traced forward to the spec and the tasks, and each task back to a requirement.
  3. SlicesThe work is cut into small slices. Each is one small change by one agent that starts fresh.
  4. EvidenceEach slice leaves proof in an evidence folder: what was run and what was seen.
  5. Blind reviewA reviewer sees only the result and the brief. It never sees the spec or the plan.

02

Who does what

One person decides. Agents from three families do the rest.

  • The personWrites the brief, makes the decisions, and accepts or rejects the result.
  • The orchestrating agentSplits the brief into requirements, writes the spec and plan, hands out slices and keeps the state files.
  • Slice agentsEach takes one small change, starts with no memory of the others, and writes its evidence.
  • The blind reviewerChecks the result against the brief. A trailing reviewer also checks each finished slice at three screen widths and answers PASS, FIXED or BLOCKED. It never publishes or commits.3

The same method works with agents from different AI providers. In one runtime, choosing a different model for each sub-agent was not available, so every sub-agent ran on the same setting.3

03

Rules

The limits recorded in the lab’s October 2026 workflow notes.

  • Trace both waysEvery requirement maps to the spec and to tasks, and every task maps back to a requirement.
  • Keep state in filesThe run moves through phases 0 to 8. Its state lives in six files: brief, spec, plan, waves, acceptance and dashboard.
  • Small runsAt most 16 tasks in a run and at most 2 at the same time.
  • One task, one fresh agentOne small change per agent, which starts with a clean context.
  • Blind acceptanceThe acceptor sees only the result and the brief. There are at most 5 rounds of rework.
  • Live means probedA claim that something is live must come from a probe of the running thing.
  • One owner per fileEach shared file has one owner at a time, so two agents never edit it together.
Report only observed results. Distinguish edited, tested, committed, pushed and deployed.Lab agent instructions, rule 7 (source 4)

04

Receipts

What the working folders held when the lab counted them on 2 Oct 2026.

Projects with the method’s folder
4
Requirements in the largest brief
88
Evidence files, in 8 folders
98
Dated acceptance records
10

Acceptance records, 24–28 Sep 2026

DateScopeResult as recorded
24 SepRequirements R1–R1312 of 13 (92%): 11 closed, 2 partial
25 SepR14–R152 of 2 in the verifiable scope
25 SepLight blind pass on version 189 of 9 applicable. The agent’s own count had been 89%
26 SepCarry-over table, version 2010 of 10
26 SepR26–R315 of 5 checkable
27 SepR603 closed, 3 partial, 0 open, of 6
27 SepR63–R678 of 8. A preservation snapshot, not a blind review
27 SepR68, the close approachAbout 75%
28 SepR69–R72About 90%
28 SepR75No percentage given

The records are not all blind passes, and their scores are not comparable: some count closed requirements, some count checks, one is a snapshot. They are listed as recorded.2

05

What did not work

Five things the records say went wrong, and what changed.

  • Choosing a model per sub-agentNot available in one runtime, so a run could not route different work to different models.
  • A stale cached pageA recheck ordered after the work was finished found a real bug. The changed page had gone out under an old service-worker cache, so returning users would have kept the old version. The cache was bumped and the page redeployed.
  • A silent revertA build on a remote machine fetched the pushed code, so a fix that existed only locally was dropped and the build failed again. The rule now is to push before any remote build that fetches from the repository.
  • A false alarmA note about a page that never existed came from misreading a screenshot.
  • An arrival that read as a flat zoomThe fix is written. It waits for a GPU to check it on.

06

Further reading

  • Library recordsGuides on agent instructions, memory and research architecture are in the Library.

07

Sources

  1. Lab working notes, 2026: the method description (phases, state files, task limits, blind acceptance), kept in each project’s working folder.
  2. Lab working notes, 2026: acceptance records, 24–28 Sep 2026, for the largest project (10 dated records; its brief runs to R88).
  3. Lab working notes, 2026: the reviewer protocol (PASS, FIXED or BLOCKED for each finished slice) and the final report of one multi-agent run.
  4. Lab agent instructions, rule 7: “Report only observed results. Distinguish edited, tested, committed, pushed and deployed.”
  5. Counts of working folders and evidence files, made on 2 Oct 2026 by listing the folders.

Facts as of 2 Oct 2026.