The older Part 4 described a visual editor that generated Python Sentinels. That editor is outside the current supported implementation. Today, extending Starlight means registering a domain agent with a small JavaScript contract and, where needed, a verifier.
Keep selection free of side effects
An agent declares its name and capabilities, then uses canHandle(intent) to indicate whether it can attempt a goal. Selection may happen without execution, so this check should not write files, reserve inventory, send messages, or call a paid model. Perform work only inside the execution path.
Make the contract easy to inspect
Part 1 includes a complete word-counter registration. For a business workflow, use the same pattern but validate input types, resource scope, and constraints before acting. Return a clear failure when the request is outside the agent’s responsibility.
Verify the outcome outside the success message
A report-writing agent should read the file back. An API agent should inspect the returned record or query its final state. A browser agent should assert the resulting application state. Merely repeating the agent’s own “completed” label does not add independent evidence.
The packaged data-report agents provide a concrete implementation to study before designing a custom verifier. Treat their code as the reference for the current hook signature and result shape.
Load local agent modules through the CLI
The CLI supports CommonJS and ESM modules. Inspect the example registrations before running a mission; the agent inventory helps expose unexpected implementations or missing capabilities.
node bin/starlight-platform.js agents --agents examples/data-report/agents.cjsA domain agent is trusted code
The platform is not a sandbox. Agents run with host permissions, and a browser or model integration can have its own network and data behavior. Review the implementation, restrict its runtime environment, and keep secrets out of mission inputs and output evidence.
- Start with one capability and a deterministic fixture.
- Exercise unsupported input and failed verification.
- Test cooperative cancellation during real work.
- Use separate ownership or locks for shared mutable resources.
- Add model planning only where the task actually needs it.
This is a narrower extension story than a no-code marketplace, but it gives each agent a contract that can be read, tested, and maintained.
Continue the series
- Starlight Part 1: From Browser Automation to a General Agent Platform
- Starlight Part 2: Reading Execution Evidence Instead of Trusting a Green Badge
- Starlight Part 3: Bounded Missions, Cancellation, and Safe Retry Decisions
- Starlight Part 5: The Core Protocol and Authenticated Remote Agents
Reviewed implementation and setup
