Bounded context
Give the agent only the messages, references, history, and fields needed for its current role.
Agent Inbox
Start with one useful agent task. Keep authority with the business.
Choose one repeatable task and the result you want. Agent Inbox gives the agent only the context and capability needed to perform that role, while company-defined permissions, limits, approvals, and result contracts determine what it may actually do.
The agent work contract
An agent proposes. The Inbox decides what may reach the business.
Agents do not need unrestricted access to become useful. They need a clear role, the right context, approved capabilities, and a predictable path for exceptions.
- 01
Propose
The agent submits intent, required context, capability, and expected result as a structured message.
- 02
Validate
The Inbox checks schema, purpose, permissions, current state, limits, and policy.
- 03
Approve
Exceptions or material actions reach the accountable expert when judgment is required.
- 04
Execute
An approved tool or Companion performs the operation and returns the outcome and evidence.
How an Active Inbox works
Start with the objective. Add what the work needs.
You do not need to model the complete operation before you begin. The first Playbook can use what is already known, then evolve through validated, controlled changes.
01
Work in motion
Something happens
A request, change, question, document, or signal reaches the Inbox.
02
Active objective
The Inbox makes sense of it
It keeps the relevant information, current state, and objective together.
03
Company authority
The right response follows
People, systems, or AI act only when your conditions allow it.
Context in
Agent Inbox: Agent requests, tool calls, proposed actions, relevant business state, prior outcomes, and role-specific evidence.
Governed outcome
The agent investigates, proposes, escalates, or performs only the operations permitted for its current role and message.
What the Inbox contributes
Useful autonomy without open access.
Approved capabilities
Choose the systems, APIs, MCP services, and Companions the agent may invoke.
Expert exceptions
Route uncertainty, policy exceptions, and material decisions to the right accountable person.
Operational evidence
Keep the request, context, policy decision, action, result, and override in one traceable path.