Notes for Claude Code

This directory is a workshop. The person you are working with is here to learn Claude Managed Agents. You write everything; they paste one prompt per part and then talk to the agent you built. Work one part at a time.

What they will say

People type these in their own words. Match the intent, not the wording:

One part per request, then stop. Do not start the next part, even if you can see its stubs and it looks easy. Leave the later stubs raising NotImplementedError: the dashboard uses them to show which parts are on. The only exception is an explicit request for all of it, such as "build everything".

Do the parts in order. If an earlier part is still a stub, do that one first and say so. They will not ask you to run the check: always run it yourself with --part N and keep going until it passes. Then stop, say what to try in the dashboard, and tell them what to type for the next part.

The job

agent.py has stub functions that raise NotImplementedError, grouped into four parts. When asked to do a part:

1. Read the stub comments for that part. They give the exact call shapes and the values you cannot look up (model constants, mount paths, the beta header). Use them as written.
2. Implement only that part's functions, plus the small additions the comments name in earlier functions (the roster in setup_agent(), the memory resources in start_session(), the second skill).
3. Run python check.py --part N. It drives the live API and prints PASS or FAIL per part. Fix what fails and run it again.
4. Explain what you did in three short bullets: what changed, the Managed Agents concept it uses, and where to see it in the dashboard.

Rules

• Edit agent.py only, plus the files under Optional changes when asked. provided.py, state.py, check.py and cleanup.py are fixed.
• Do not open solutions/. It is the answer key. Work from the stub comments and the installed anthropic package.
• Create cloud resources only through state.sync(...), with state.name(...) in every name= and state.owner() in every metadata=. A whole room may share one API key, and this is what keeps people's resources apart.
• Never print, log or paste the API key. Never put it in a prompt, an event or a resource.
• Do not delete or archive anything through the API yourself. python cleanup.py does that safely.
• Use the platform's exact names when you explain things: agents, environments, sessions, events, skills, custom tools, multiagent, threads, memory stores, session outputs.

Optional changes

The person does not edit files by hand. After a part passes they may ask you for a change to one of these. Make it for them, in the file named:

Changes take effect on the next dashboard rerun or check.py run: a changed prompt becomes a new agent version, a changed skill folder becomes a new skill version, and a changed memory-seed/ becomes a freshly seeded store. Sessions already running keep what they started with, so start a new session to see a change.

Useful facts

• Run the dashboard with streamlit run app.py. The chat panel shows which parts are on.
• The log is mounted in the container at /mnt/session/uploads/app.log.
• Custom tools run on this machine in handle_tool(). A sub-agent's custom tool call arrives on the session stream with session_thread_id set; answer it like any other.
• Memory stores attach when a session is created, not afterwards.
• Files the agent writes under /mnt/session/outputs/ appear in files.list(scope_id=session_id) a few seconds after the turn ends.
• Tested with anthropic==0.116.0. If a call shape in a stub comment disagrees with a newer SDK, say so instead of guessing.