FAQ (AI-curious engineers)

Short answers for folks who like agents, MCP, and shipping real integrations. Deeper detail lives on the linked pages.

1. How do agents actually call this?

MCP (stdio server in @acoyfellow/lab-mcp) or raw HTTP: GET /lab/catalog for capability and execute metadata, then POST /run, /run/chain, etc. See Agent integration and HTTP API. A minimal loop is: discover paths from the catalog, run, read resultId, then fetch GET /results/:id.json. Humans can open GET /results/:id in the app viewer. Successful runs include full step data; failed or aborted runs may include partial-or-empty step detail.

2. How tight is the sandbox for LLM-generated or user-supplied JS?

Guest code runs in Worker Loader isolates with an explicit capability set per run (and per chain step). No capability ⇒ no shim for KV, spawn, Workers AI, R2, D1, DO, containers. /run/generate asks Workers AI to emit a guest body, then runs it under the capabilities you passed—same isolation model as hand-written code. See Security, Capabilities, and Architecture.

3. What’s the observability story for agent runs?

Every persisted run returns a resultId. The receipt document contains the request shape, outcome, timing, and per-step data for successful chains. Agents should read /results/:id.json. Humans can open /results/:id as the viewer over that same receipt. It’s a shareable artifact, not a full “agent session” product—you correlate tool calls to receipts in your own control plane. See Receipt schema and Failures & step data (including chain failure behavior).

4. Cost and limits—what breaks first?

Platform CPU/time and Workers limits apply; Lab doesn’t add a separate chain step cap in code. Spawn depth is bounded; R2 invoke caps list/size in the Worker. See Limits. For dollars-on-the-bill precision, use Cloudflare’s published pricing and your account—not the receipt’s wall ms alone.

5. Why edge instead of a Python sidecar or local sandbox?

Fit: short, chainable, least-privilege guest JS next to KV / AI / R2 / … with receipts. Poor fit: long CPU-heavy batch jobs or arbitrary host access—see When to use Lab and Architecture.