autonomous-loops

name	autonomous-loops
description	Patterns for autonomous agent loops — sequential pipelines, parallel generation, continuous PR loops, and de-sloppify passes. Use when setting up automated workflows, CI/CD agent loops, batch processing with Claude, or multi-agent orchestration. Covers loop architecture from simple sequential to RFC-driven DAG orchestration.

Autonomous Loop Patterns

Six loop architectures of increasing sophistication. Pick the simplest one that fits.

1. Sequential Pipeline

Each step is isolated (fresh context). Order matters, exit codes propagate.

claude -p "Implement auth module per spec in PLAN.md" && \
claude -p "Review changes, fix issues, run tests" && \
claude -p "Write commit message, commit"

Use when: Simple linear tasks, each step is independent.

2. De-Sloppify Pipeline

Two focused agents beat one constrained agent.

claude -p "Implement feature X with full TDD"
claude -p "Review all changes. Remove: tests of language behavior, console.logs, dead code, over-engineering. Run test suite."

Use when: Any implementation task. Always follow implementation with cleanup.

3. Parallel Generation

Spawn N agents with different creative directions. Each gets an iteration number to prevent duplicates.

for i in $(seq 1 5); do
  claude -p "Iteration $i: Generate solution for X. Be creative, try approach $i." &
done
wait
claude -p "Review all 5 solutions in output/. Pick the best, explain why."

Use when: Creative/exploratory tasks where multiple approaches might work.

4. Continuous PR Loop

Agent runs in a loop: create PR → wait for CI → fix failures → repeat.

Key elements:

• SHARED_TASK_NOTES.md bridges context gaps between iterations
• Auto-fix failing CI checks
• Configurable max-runs, max-cost, completion signals
• Merge queue with conflict recovery

Use when: Automated PR workflows, CI-driven development.

5. Spec-Driven DAG Orchestration

Most sophisticated: decompose spec into tiered work units with dependency DAG.

• Each unit runs through quality pipeline in isolated worktree
• Separate context windows per stage (reviewer ≠ author)
• Merge queue with eviction recovery

Use when: Large features with multiple interdependent components.

Decision Flow

Need strict CI/PR control?      → Continuous PR loop
Need RFC decomposition?          → DAG orchestration
Need exploratory parallel gen?   → Parallel generation
Simple linear task?              → Sequential pipeline
Any implementation task?         → Always add de-sloppify pass

Failure Modes to Watch

• Loop churn without measurable progress → freeze, audit, reduce scope
• Repeated retries with same root cause → investigate, don't retry
• Merge queue stalls → check for dependency cycles
• Cost drift from unbounded escalation → set --max-budget-usd