autonomous-loops

Autonomous Loops

A catalog of architectures for running a coding agent repeatedly with little or no human intervention between iterations, from a simple scripted sequence of non-interactive invocations up to a full RFC-driven multi-agent DAG. Pick the simplest pattern that fits the task — most autonomous work needs pattern 1 or 2, not the most sophisticated one on this list.

Pattern spectrum

1. Sequential pipeline

The simplest loop: a script that invokes the agent non-interactively, one focused step at a time, each building on the filesystem state the previous step left behind.

#!/bin/bash
set -e

# 1. Implement, TDD
claude -p "Read docs/auth-spec.md. Implement OAuth2 login in internal/auth/. Write tests first."

# 2. Cleanup pass (see De-sloppify below) — separate context, separate concern
claude -p "Review files changed by the last commit. Remove tests that verify language/framework behavior rather than business logic. Keep real logic tests. Run go test ./... after."

# 3. Verify
claude -p "Run go build, go vet, and go test ./... . Fix any failures. Do not add new features."

# 4. Commit
claude -p "Create a conventional commit for the staged changes."

Design principles:

• Each step is isolated. A fresh invocation per step means no context bleed from an earlier, now-irrelevant step.
• Order matters and is explicit. The script is the source of truth for sequencing, not the model's judgment about what to do next.
• Avoid negative instructions in the implementer step ("don't write pointless tests") — telling a model what not to do inside a step that also has to do a lot of positive work makes it hesitant across the board. Add a separate cleanup step instead (see below).
• Exit codes propagate. set -e stops the pipeline on the first failure rather than compounding it into the next step.