Agent operating system

How I run design work with agents, and how I keep them honest

May 2026 to present

  • Personal system
Role
Sole designer and builder
What it is
An agent system for my design work, with self-checking jobs and separate verifier agents
Result
Over six months: 381 sessions, 651 subagents, and a little over 35,000 tool calls.
Hooks: handoff, skill trigger, auditMeMain sessionorchestratesSubagentsdo the workAgent jobsrun on a loopPending reviewone-way doorsVerifierseparate agentCommitMemory and wiki on diskdirectsdelegateschecked bycommitsparkswaits for mereads and writesHooks: handoff, skill trigger, auditMeMain sessionorchestratesSubagentsdo the workAgent jobsrun on a loopPending reviewVerifierseparate agentCommitMemory and wiki on diskdirectsdelegatescheckedcommitsparksasks mereads and writes
Each agent job passes a verifier first. Then a reversible change is committed, and anything I cannot undo is parked in pending review for me.

Most of my design work now runs through AI agents. A separate verifier agent checks every job that changes something, and anything I can’t undo waits for my approval.

The verifier that passed a broken layout

Beforeshellempty gapcontentPASS 94AftershellcontentPASS 85Beforeshellempty gapcontentPASS 94AftershellcontentPASS 85
Redrawn, not the real test file. Left: the layout with a gap that a verifier passed at 94. Right: content inside the shell, scored 85 by the rubric that reads real dimensions.

In August a verifier gave a Figma build 94 out of 100. The layout was broken. I changed the rubric that day to read real dimensions, and the re-run scored 85. Since then every verifier I write has structural checks. “No error” isn’t a pass.

Six months of activity

From May to October 2026: 381 sessions, 651 subagents, and a little over 35,000 tool calls.

May5,477from May 5June5,011July409sabbaticalAugust9,055September6,725October8,598to Oct 29Mayfrom May 55,477June5,011Julysabbatical409August9,055September6,725Octoberto Oct 298,598
Tool calls, subagents, and browser checks per month, May to October 2026
MonthTool callsSubagentsBrowser checks
May (from May 5)5,47710286
June5,011109118
July (sabbatical)4092310
August9,055170108
September6,725111207
October (to Oct 29)8,598136215
Tool calls per month, May 5 to October 29, 2026. July was a sabbatical, so its bar is hollow.

How it was built

I run six recurring agent jobs. Each keeps its state in a file on disk.

loops/dream-loop.md

loop: dream-loop
type: heartbeat-with-threshold
schedule: self-paced (in-session)
success_criteria: >
  Either (a) thresholds unmet and the cycle noops cheaply, or (b) a scoped dream
  pass ran, a verifier subagent cleared the diff, reversible changes were committed,
  and any gated items are parked in Pending review.
status: active
last_full_run: 2026-11-05
memory_count_at_last_run: 124
memory_count_gate: 135
...
- 2026-11-05 — ... Move plan rows 1, 2, 3, 5 applied, verifier 8.0,
  behavior test failed (66 vs 53), all reverted. ...
The state file of the job that keeps my agent memory clean, and a line from its run log. The verifier passed the change, a behavior test did not, and the job undid its own work. Shortened. From my home setup.

knowledge/verifier-personas/README.md

2. Walk the must-check list. Answer every ID with pass, fail, or n/a, plus
   evidence (a quote, a selector, a file and line, a screenshot path). ...
5. Caps. One open Blocker caps the overall score at 6.0. One open Major caps it
   at 8.0. ...
7. Stay independent. Do not read the builder's self-assessment or earlier
   verdicts, unless this is a fix round. ...
Three of the ten rules that every verifier agent gets. Evidence for every row, a cap on the score while a defect is open, and no look at the builder's own grade. Shortened. From my home setup.

Hooks hand a long session to a fresh one with the state on disk. Git rules: a branch for everything, no push without my go-ahead. More in How I work with AI and my loop engineering article.