# The agent swarm

Tommy Tai builds most of his work with a swarm of AI workers from several model vendors. The swarm's coordinator is Sigrún, an AI. Its public operating rules are the Sigrún sliver kit: [sliver/SKILL.md](sliver/SKILL.md) and [sliver/doctrine.md](sliver/doctrine.md).

## How it works

- One coordinator turns the operator's direction into small task cards.
- Stateless workers from an allow-list of vendors take one card each and return a result with a receipt (a commit, a file, a test run).
- Work goes through several rounds: plan, do, study, act. The operator sees only champions: results that won several rounds of checks, not first drafts (operator's standing rule, 2026-10-30).
- A result counts only after a checker from a different model vendor has run it. Deterministic checks (tests, exact verifiers, browser probes) come first.
- The checkers are tested with planted bugs in three tiers. Rule of thumb: an 80 to 95% catch rate; 100% means the planted bugs are too easy.
- People keep the irreversible steps: money, messages to people, merges, deploys, deletes.

## Numbers

- One documented catch: a swap test printed "SWAPPED" for all five parts it was meant to exercise, but a test setting made a stub run instead of the real code; a gate run by a different model family read the code and rejected it (gate note, 2026-10-29, private working repository).

## Not claimed

- That a planted-bug catch rate measures how many real bugs get caught.
- That the approach works beyond Tommy Tai's own projects. That is still a thesis.

## Writing

- Swarm log post, 2026-10-30: "Plant bugs your AI agents should catch", https://tommytai.dev/blog/2026-10-plant-bugs/
- Human page: https://tommytai.dev/work/agent-swarm/