The agent page

Every agent has one page at /dashboard/agents/:id. Its sections are:

• Overview: status, role, learning period and risk at a glance.
• Connections: the systems the agent reaches and the tools it may use on each.
• Data: data protection and attack settings for this agent.
• Approvals: the approval rules that make a person decide before a call runs.
• Limits: allowed models, request rate and budget.
• Activity: the activity ledger of what the agent read, changed, ran and sent.
• Exposure: what the agent could reach if it misbehaved.

Approvals, Data and Limits also have an organization view with a per-agent table. Both views edit the same setting, so there is never a second place a per-agent value hides.

An agent reaches systems through connections, and a tool is a pair of system and tool. Every access decision comes from access rules of one shape (who, what, when, then), layered organization, then agent, then exception.

The first MCP tool call to a server outside the agent's connected set asks a person once (in monitor mode it is only recorded). Deciding that approval approves or blocks the connection.

Read and change an agent's access over the API:

GET  /api/agents/:id/access
PUT  /api/agents/:id/access
GET  /api/agents/:id/data-policy   PUT /api/agents/:id/data-policy
GET  /api/agents/:id/limits        PUT /api/agents/:id/limits
GET  /api/agents/access/replay     # what a change would have decided differently

Limits combine across layers: models narrow, the lower rate wins and budgets add up. A call over a limit is refused with 403 (model), 429 (rate) or 402 (budget).