Claude Managed Agents on Render Sandboxes

Run a Claude Managed Agent's tool calls in Render Sandboxes, one fresh sandbox per session. The orchestrator comes in TypeScript and Python. Both versions behave the same and pass the same tests.

• Tutorial: Run Claude Managed Agents in Render Sandboxes deploys the orchestrator, has an agent fix a program, and verifies the result. The starter/ directory holds its helpers.
• Reference: GUIDE.md covers staging files, persistence, network rules, settings, security, and troubleshooting.
• Deploy (TypeScript): use Deploy to Render. This branch puts the TypeScript Blueprint at the repository root for the button.
• Deploy (Python): use Deploy to Render from the default branch.
• Test (TypeScript): cd orchestrator-ts && npm test.
• Test (Python): pytest tests/test_orchestrator.py, then pytest -m live tests/test_render_live.py with Render credentials, then python scripts/e2e.py with Anthropic credentials.
• Real-session checks: python scripts/validate.py runs real Claude sessions through the Python orchestrator. Set ORCHESTRATOR=ts to run them through the TypeScript one.

orchestrator-ts/      TypeScript orchestrator, its Blueprint, and offline tests
orchestrator/         Python orchestrator (Render background worker)
sandbox/              bootstrap script that runs inside each sandbox (shared by both)
scripts/              Python snapshot builder, end-to-end test, and real-session checks
starter/              the tutorial's Python and TypeScript helpers, task fixture, and verifier
examples/typescript/  create an environment and agent, and run a first session
tests/                Python offline tests and live Render tests