Loop Engineering for agents

Give Agents a Loop That Closes

Ariadne Loop turns scattered context into a loop contract: inspect, act, verify, decide.

python -m pip install git+https://github.com/zhangzeyu99-web/ariadne-loop.git
ariadne-loop quickstart --output .ariadne/quickstart
# hand agent-packet.md to your coding agent
A glowing teal loop running through an abstract architectural workspace
Inspect Act Verify Decide

For evidence-led agent work

Prompts are not enough. Keep scope, gates, reports, and human handoffs in one place.

Loop specs

Turn rough notes into inspect-act-verify-decide loops.

Browser builder

Draft snapshot JSON in the browser.

Agent packets

Hand off scope, constraints, gates, and budget.

Report validation

Require parseable evidence, not status prose.

Loop supervision

Catch drift, repeated failures, and rollback signals.

Agent recipes

Ready loops for repairs, releases, handoffs, and PR follow-ups.

Claude Code command

Run /ariadne-loop inside a project.

OpenClaw operations

Put scheduled work behind checks and stop rules.

The loop is the guardrail

Ariadne builds the maze and leaves the thread. Every pass stays observable.

1

Inspect real state

Read the real project before changing it.

2

Act small

Make one bounded change.

3

Verify evidence

Run the gate and keep the proof.

4

Decide the next turn

Stop, continue, roll back, or ask.

Start with a snapshot.

Capture state once. Hand the packet to your coding agent.

Build a loop