How an agent calls a tool
An agent cannot act on the outside world by itself. To do anything — reply in Slack, fork, merge, save its memory, or start another agent — it calls a tool on the MCP server. The server is the agent's only way out. Every call passes through it, so it is also where every rule is enforced.
The agent has a small, fixed set of tools and nothing else. Each call is checked before it runs:
- Who is calling. The server reads the agent's identity from the running process, not from the call, so a tool call cannot claim to be someone else. An unknown caller is refused.
- What it may touch. An agent may act on itself and its own descendants, never a sibling or an ancestor. Anything out of scope is refused before it runs.
- Whether a human agreed. A few actions wait for a person. See below.
A human's OK
An agent makes most decisions on its own. A few are risky enough to wait for a person: merging work back, and shutting another agent down. The MCP server holds these actions until a human agrees.
The agent proposes the action instead of running it, and a note goes to Slack. A human agrees with a reply or a ✅ reaction. Only then does the action run, and it runs once. A second ✅ does nothing.
Starting or waking an agent does not wait for approval. The message that triggered it already counts as approval.