Composing a loop with adept
A loop is a system that prompts agents so a human doesn't have to: each turn it
discovers work, hands off to an agent, verifies the result, persists state,
and reschedules. Skip any move and the loop fails a predictable way — no verification is
an agent nodding at itself; no persistence starts from scratch every morning; no schedule is
a script someone forgets to run.
adept manages two of the six parts as canonical, synced resources — skills (discovery
knowledge) and agents (the generator/evaluator pair). The rest (worktrees, connectors,
memory files, schedules) belong to the harness and repo; this skill tells you how to wire
them together.
Scaffold the composition
adept loop add ci-triage --workflow
creates:
skills/ci-triage/ — discovery skill (triage template: Read → Judge → Write → Hand off →
Stop). activation: manual on purpose: the automation invokes it by name; it must not
auto-fire in normal sessions.agents/ci-triage-reviewer.md — evaluator agent (adversarial template: assume broken,
act don't read, PASS/REJECT verdict)..github/workflows/adept-loop-ci-triage.yml — cloud cron skeleton (only with
--workflow); fill in the harness invocation + auth secret.
Then edit the scaffolds, adept agent check ci-triage-reviewer, adept sync.
The five moves, mapped
| Move |
Where it lives |
adept's part |
| Discovery |
triage skill — what to read, how to judge noise vs. actionable |
adept skill add <id> --template triage |
| Handoff |
one git worktree per finding (claude --worktree, Codex background worktree) |
skill's "Hand off" section emits the task lines |
| Verification |
evaluator agent, different instructions (ideally different model) than the maker |
adept agent add <id> --template evaluator |
| Persistence |
./state/<id>.md committed to the repo — the agent forgets, the repo does not |
skill's "Write" section owns it |
| Scheduling |
cloud cron (machine off, ≥1h interval) or local /loop/automations (machine on, minute-level) |
--workflow skeleton |
Non-negotiables (the parts that produce safety, not output)
- Independent evaluator. An agent grading its own work praises it. Keep maker and
checker separate; tune the checker skeptical; make it verify by acting (run tests, click
the page), and let a stop condition (e.g. Claude Code
/goal) be judged by a fresh model.
- The Stop section. The skill's boundary ("Never merge. Never delete. Uncertain →
./inbox/ for a human") is the one thing the loop cannot infer. Leave it out and the loop
merges with confidence it has not earned.
- Human review stays. PRs open, never auto-merge. The pause is a permanent feature, not
scaffolding to remove once trusted.
- Cap before you ship. Per-run budget, daily budget, max retries — set them before the
first unattended run, because one bug spinning idle overnight otherwise becomes a bill.
- Grow parallelism last. Prove the evaluator catches real mistakes on one finding
end-to-end before letting five agents run at once.
Four silent costs (watch for them, tell the user)
Verification debt (unreviewed output), comprehension rot (code nobody read), cognitive
surrender (nobody says "this is wrong" anymore), token blowout (retries all night). They
reinforce each other; the guard for all four is a check that can say "no" plus a human who
still reads a sample daily.
See [[authoring-adept-skills]] for the discovery skill's craft, [[authoring-adept-agents]]
for evaluator design, and [[using-adept]] for the CLI. adept loop add --help is the source
of truth for flags.
1---2name: authoring-adept-loops3description: Compose a loop — a scheduled system that discovers work, hands it to agents, verifies with an independent evaluator, persists state, and reschedules itself. Apply when the user wants automation that runs on a timer, a triage/babysitter routine, or asks about `adept loop add`.4---56# Composing a loop with adept78A loop is a system that prompts agents so a human doesn't have to: each turn it9**discovers** work, **hands off** to an agent, **verifies** the result, **persists** state,10and **reschedules**. Skip any move and the loop fails a predictable way — no verification is11an agent nodding at itself; no persistence starts from scratch every morning; no schedule is12a script someone forgets to run.1314adept manages two of the six parts as canonical, synced resources — **skills** (discovery15knowledge) and **agents** (the generator/evaluator pair). The rest (worktrees, connectors,16memory files, schedules) belong to the harness and repo; this skill tells you how to wire17them together.1819## Scaffold the composition2021```bash22adept loop add ci-triage --workflow23```2425creates:2627- `skills/ci-triage/` — discovery skill (triage template: Read → Judge → Write → Hand off →28 Stop). `activation: manual` on purpose: the *automation* invokes it by name; it must not29 auto-fire in normal sessions.30- `agents/ci-triage-reviewer.md` — evaluator agent (adversarial template: assume broken,31 act don't read, PASS/REJECT verdict).32- `.github/workflows/adept-loop-ci-triage.yml` — cloud cron skeleton (only with33 `--workflow`); fill in the harness invocation + auth secret.3435Then edit the scaffolds, `adept agent check ci-triage-reviewer`, `adept sync`.3637## The five moves, mapped3839| Move | Where it lives | adept's part |40|---|---|---|41| Discovery | triage **skill** — what to read, how to judge noise vs. actionable | `adept skill add <id> --template triage` |42| Handoff | one git worktree per finding (`claude --worktree`, Codex background worktree) | skill's "Hand off" section emits the task lines |43| Verification | evaluator **agent**, different instructions (ideally different model) than the maker | `adept agent add <id> --template evaluator` |44| Persistence | `./state/<id>.md` committed to the repo — the agent forgets, the repo does not | skill's "Write" section owns it |45| Scheduling | cloud cron (machine off, ≥1h interval) or local `/loop`/automations (machine on, minute-level) | `--workflow` skeleton |4647## Non-negotiables (the parts that produce safety, not output)4849- **Independent evaluator.** An agent grading its own work praises it. Keep maker and50 checker separate; tune the checker skeptical; make it verify by *acting* (run tests, click51 the page), and let a stop condition (e.g. Claude Code `/goal`) be judged by a fresh model.52- **The Stop section.** The skill's boundary ("Never merge. Never delete. Uncertain →53 `./inbox/` for a human") is the one thing the loop cannot infer. Leave it out and the loop54 merges with confidence it has not earned.55- **Human review stays.** PRs open, never auto-merge. The pause is a permanent feature, not56 scaffolding to remove once trusted.57- **Cap before you ship.** Per-run budget, daily budget, max retries — set them before the58 first unattended run, because one bug spinning idle overnight otherwise becomes a bill.59- **Grow parallelism last.** Prove the evaluator catches real mistakes on one finding60 end-to-end before letting five agents run at once.6162## Four silent costs (watch for them, tell the user)6364Verification debt (unreviewed output), comprehension rot (code nobody read), cognitive65surrender (nobody says "this is wrong" anymore), token blowout (retries all night). They66reinforce each other; the guard for all four is a check that can say "no" plus a human who67still reads a sample daily.6869See [[authoring-adept-skills]] for the discovery skill's craft, [[authoring-adept-agents]]70for evaluator design, and [[using-adept]] for the CLI. `adept loop add --help` is the source71of truth for flags.
itaywol