Cristhianzl Running agent loops
Patterns for driving work across many steps or iterations — possibly unattended — without the run drifting, looping, or shipping slop. The core trade is isolation vs. continuity: a fresh context per step avoids bleed, but you must deliberately carry forward what matters.
Read first (always)
List learnings/ and read anything relevant. Project-specific loop conventions, budgets, and stop signals live there and override this file.
Principles (apply to every loop)
- Fresh context per step beats one long context. Each
claude -p/ subagent call starts clean — no bleed, no drift. The cost: it forgets. Bridge it deliberately (below). - The reviewer is never the author. Put review/verify in a separate step/context so it isn't anchored to the implementation.
- Two focused passes beat one constrained pass. Don't pile negative instructions onto the implementer — add a separate de-sloppify pass (below). Quality from constraints decays; quality from a dedicated cleanup step doesn't.
- Every loop has an explicit stop condition. Bound it by max-runs, max-cost, max-duration, or a completion signal — never "until it feels done".
The patterns (pick the simplest that fits)
| Pattern | Use when | Shape |
|---|---|---|
| Sequential pipeline | A known series of steps on one unit of work | implement → de-sloppify → verify → commit, each a fresh context |
| PR loop | Iterate a branch toward green | branch → implement → review → fix → run checks → repeat until checks pass or budget hits |
| Parallel fan-out | N independent variations/units of the same spec | an orchestrator assigns each agent a distinct direction + index (don't rely on agents to self-differentiate); run in waves of 3–5 |
| RFC → DAG | A large feature with dependencies | decompose into work units + a dependency DAG; run each in an isolated worktree; land via a merge queue; deeper review tier for riskier units |
The cross-iteration context bridge
Independent steps lose memory. Persist a single SHARED_TASK_NOTES.md (or the project's plan file): the agent reads it at the start of each iteration and updates it at the end — current objective, what's done, what's next, decisions, and gotchas. This is what turns isolated calls into a coherent run.
The de-sloppify pass
After each implement step, run a dedicated cleanup pass (separate agent/context) whose only job is to remove slop: dead code, stray comments, debug prints, half-finished branches, inconsistent naming, over-engineering. Keep it separate from the implementer so neither job is diluted.
Safety before you launch an unattended loop
- Baseline tests pass before iteration 1.
- An explicit stop condition and a budget (runs / cost / time).
- Rollback ready: work on a branch or in a worktree, commit per green step.
- Checkpoints so you can diff now-vs-then.
- Escalate to the human when: no progress across two consecutive checkpoints, the same error/stack repeats, the diff drifts off-objective, or the budget is exceeded. Stalling silently is the failure mode.
If a loop is wedged (looping, repeating an identical failed action, burning budget), switch to skills/debugging-agent-runs.
Capture a learning
If you hit a loop convention, stop-signal, or trap not covered here, append a learnings/YYYY-MM-DD-slug.md (or use /learn).
See also
/dual-reviewcommand — the adversarial two-reviewer convergence step for a PR loop.skills/debugging-agent-runs— recover a wedged run.skills/reviewing-code,skills/writing-pull-requests— the review/PR steps a loop calls.