AWS and Agents quickstart#

Rat Things keeps its API, harness, session state and execution in your AWS account. Use the deployment guide to configure the Terraform module, model credentials, ARM64 runtime image and Agents HTTP service. Package Lambda artifacts before Terraform validation or deployment. Deployment creates AWS resources; ordinary local checks do not provision MicroVMs.

Automated setup#

The setup helper deploys the AWS backend, creates an Agent with the selected model, and completes two Turns in one Session. It checks the saved assistant Items, deletes the disposable Session to stop its harness, and leaves the Agent available for new work. Setup provisions AWS resources and invokes the model.

npm run quickstart:aws -- --model YOUR_ADMITTED_MODEL

For ChatGPT authentication, setup requires explicit consent before copying the device's renewable Codex credentials into your AWS Secrets Manager. --model selects a model admitted by that credential and deployment. Use --auth bedrock with a configured Bedrock model for AWS model authentication instead.

--dry-run describes setup without deploying. Existing-deployment status, credential synchronization and teardown commands retain the deployment's saved AWS identity. The local deterministic smoke test is npm run smoke:local; a mock driver cannot prove an Agents deployment.

The helper uses the Lambda URL transport when a dedicated Agents HTTP service is not configured. Large uploads, long-lived SSE and executor connections need the public transport deployment.

Connect an installed deployment#

Set the Agents endpoint from agents_api_base_url and select an admitted model:

export RAT_THINGS_AGENTS_API_URL="https://your-agents-endpoint/v1"
export AWS_REGION="us-west-2"
export RAT_THINGS_MODEL="your-configured-model"

npm run rat-things -- sessions create --model "$RAT_THINGS_MODEL" \
  --input "Explain Agents, Sessions and Turns." --stream

The CLI obtains an owner-scoped API key using the workstation's AWS identity. The identity must be allowed to invoke the deployment's token issuer. Model requests use the deployment's configured provider and credentials.

For reusable work, save an Agent request containing model, name, instructions and declared tools, then create it:

rat-things agents create --file agent.json
rat-things sessions create --agent-id agent_example --input "Begin the review."
rat-things sessions turns sess_example
rat-things sessions items sess_example
rat-things sessions send sess_example --input "Also include the rollout plan."

The default CLI Session environment is none. Use a standard Session JSON request with an environment template when commands or a filesystem are required. Agent configuration is snapshotted when the Session is created.

Use the SDK guide for complete resource examples and schedules and provider bindings for automated input. Set RAT_THINGS_API_URL to the separate control endpoint for schedule and connection installation commands.