How to keep Codex running after you close your laptop#

Run the agent on remote compute and retain its inputs and results outside your laptop. Rat Things runs the API, harness, storage and execution in your AWS account. A Session identifies the work; Turns record each execution and Items retain its messages and tool activity.

Choose where work runs#

A terminal multiplexer can survive closing a terminal on a machine that stays awake. A remote worker lets work continue when your laptop sleeps. Durable storage also preserves saved history and files if the worker is lost, but it cannot recreate an arbitrary running process.

Rat Things supports isolated Lambda MicroVM execution and an opt-in dedicated EC2 Session worker. A connected EC2 worker can retain processes between Turns without being subject to the MicroVM's eight-hour maximum lifetime. Retained state and a live environment are separate properties: inspect Session and environment status before assuming work can continue.

Submit a Session#

Configure RAT_THINGS_AGENTS_API_URL, AWS_REGION and your AWS credentials for the deployment. Replace the model placeholder with a model admitted by that deployment:

npm run rat-things -- sessions create \
  --model YOUR_ADMITTED_MODEL \
  --input "Draft a release-readiness checklist with rollback steps."

This example uses environment: {"type":"none"} and needs no repository checkout. Save the returned Session ID before disconnecting. Later, replace sess_example with that ID to inspect and continue the work:

npm run rat-things -- sessions get sess_example
npm run rat-things -- sessions turns sess_example
npm run rat-things -- sessions items sess_example
npm run rat-things -- sessions send sess_example \
  --idempotency-key release-review-2 \
  --input "Refine the checklist into a concise go/no-go review."

A repeated input with the same idempotency key and body reuses its accepted receipt. To work with commands and files, supply a standard Session creation JSON file with a managed or self-hosted environment using --file request.json. See Agents API for those request shapes.

Retain results and control authority#

Encrypted S3 retains full content and native checkpoints; DynamoDB holds resource revisions and coordination fences. Managed output files under /workspace/outputs are saved as immutable Session artifacts when Turns finish. Download them before deleting the Session. Saved Items and artifacts survive environment expiration.

The Agent configuration, declared tools, environment policy and credential grants bound the work. A denied operation cannot open a human-approval path that widens that authority. Notification connections do not automatically grant tools to the agent. Read the capability envelope.

Close work deliberately#

Use sessions cancel sess_example to stop active work while retaining the Session. Use sessions delete sess_example when it is no longer needed; deletion removes it from the public API and closes its private harness. A connected worker can incur compute costs while waiting for another Turn.

Current boundaries#

Native checkpoints and saved workspace bytes support recovery after execution loss. They do not guarantee that background processes, sockets or unacknowledged external writes survive replacement. See Session durability and the durable-state guide.

Try the narrow path#

The AWS quickstart creates a disposable deployment, runs two root Turns in one Session, and provides teardown. Use an account and credentials intended for that work.