Mission · Set limits
Track what your agent does — and draw clear limits
Your agent is running. To finish, you learn two things: to read and interpret the state of an order — and to set durable limits on which Tools your agent may use at all. That's how you stay in control even when you're not sitting next to every single call.
What you'll be able to do
- Read an order's status and interpret the stages
prepared → approval_requested → …. - Durably allow your agent only the Tools you want — in the dialog and with
/permissions.
Read the status — get_order_status
You don't have to memorize order numbers. Ask your agent in plain language, and it calls
get_order_status (only after your permission, of course) and translates the
answer. The platform moves an order through fixed stages — the names are in English, because
that's how they appear in the app and in the log too:
prepared— prepared, waiting for you. Badge "Vorbereitet".approval_requested— the Freigabe was requested and is in your Messenger.pending_approval— waiting for your approval. Badge "Genehmigung ausstehend".rejected— declined by you. Badge "Abgelehnt". (That's how your practice order from Lesson 3 ended.)
The history shows the path with no gaps — e.g.
prepared → approval_requested → rejected. So you can read back at any time what
happened to your agent's order, and who moved it when.
Set limits — only these Tools
In the Agentowner course you set Auftrags-Limits (order limits) for your agent. Here you draw the line at the build level: which Tools it may call at all. You already know your instrument for that — the permission dialog:
- Allow once:
1. Yes— for this single call only. - Allow durably:
2. Yes, and don't ask again …— Claude Code remembers this Tool for your current project. Step by step you build a whitelist of exactly the Tools you want. - Decline:
3. No— and the Tool stays out.
With the command /permissions you see at any time what you've allowed, and can
revoke it. The decisive protection: the dialog always names the exact Tool name. If
your agent unexpectedly reaches for mcp__lys__prepare_orders (the batch variant)
or even mcp__lys__task_execute, you see it — and say No.
Pin it down durably (optional)
To anchor the whitelist for good, add the allowed Tools to the settings file
settings.json under permissions.allow — for example
mcp__lys__search_skillanbieter and mcp__lys__prepare_order. For
the course the dialog is entirely enough; this is just the way to hold it across projects.
Deliberately left out
task_execute, auto_approve and the ⏵⏵ auto mode skip
your Freigabe. That's exactly why you don't allow them in this course: they
contradict the through-line that you keep the last word.
When the platform draws limits
Two errors you'll see sooner or later. Claude Code shows both directly in the chat when a tool call fails:
- Authentication: the Token is missing or wrong — the access won't let your agent in (see Lesson 1). Check the connection with
/mcp. - Too many requests: the access limits how often your agent may call in a short time. When you hit that limit, waiting a moment and retrying helps. The exact limits are in the platform's MCP docs.
Keep learning
You've built an AI agent in the terminal, had it search and order, and set it limits — never giving up control. Two sources take you deeper:
- MCP Integration — endpoint, all Tools, Token scopes and the platform's limits.
- docs.claude.com/claude-code — MCP servers, permissions and modes in Claude Code itself.
Your turn
Test yourself — retrieving from memory is what builds real retention.
Knowledge check
How do you durably allow your agent only certain Tools?
With 2. Yes, and don't ask again Claude Code remembers a Tool durably for the
project, and /permissions shows and revokes what's allowed — so your whitelist
grows. Auto mode does the opposite: it removes your control. "Every Tool, always" would be
exactly the limitlessness you avoid here.
