autonomous-loops

Use the claude -p flag to run Claude Code non-interactively with a prompt, then chain calls in a shell script with set -e so each step builds on the previous one's filesystem state. Add exit conditions like max-runs, max-cost, or a completion signal to prevent infinite loops.

A sequential pipeline chains claude -p calls for scripted, single-focus tasks, while the Continuous Claude PR loop creates branches, opens PRs, waits for CI checks, auto-fixes failures, and merges across many iterations. Use the PR loop for multi-day iterative projects and pipelines for daily scripted steps.

Each claude -p call starts with a fresh context window, so bridge iterations with a SHARED_TASK_NOTES.md file that Claude reads at the start and updates at the end of each iteration. Filesystem state and context files passed via prompts also work.

Use DAG orchestration when you have a written RFC, multiple interdependent work units, and a need for parallel implementation with merge coordination. For single-file changes or quick iteration, a sequential pipeline or NanoClaw session is simpler and sufficient.

LLMs instructed to do TDD often add tests for language features and overly defensive checks. Instead of negative instructions in the implementer prompt, add a separate de-sloppify pass where a fresh agent removes test slop while keeping business logic tests.

Assign each agent a unique work unit with minimal file overlap, run units in isolated worktrees, and land them through a merge queue that rebases onto main and runs tests. Evicted units re-enter with full conflict context so the next attempt resolves the collision.