FOR AGENT BUILDERS · DELEGATED AUTHORITY
Agents that help.
Authority stays yours.
The future is not an assistant with unlimited access. It is useful agents working from task-specific context, with authority borrowed for a purpose and returned to people when a decision matters.
A clear answer
How should an agent receive delegated authority?
The proposed model gives an agent a task, a defined scope, and an understandable activity trail. Consequential actions should return to a person for approval; broad orchestration remains future work.
See implemented and proposed layersVISION / PROPOSED · LOCAL SIMULATION
Borrow authority for one task.
Choose a few records for a fictional studio review. Then see how a future agent could prepare within a defined scope and return a consequential action to you.
Draft a recap from selected review context.
YOU DECIDESending or changing a calendar stays with a person.
MCP is an early integration surface; broad delegated orchestration is not yet available.
“What should I have ready for the studio review?”
The review timing and studio agenda are included; the personal reminder stays out. Personal reminder enters only at the far-right setting.
In this task 2
Held outside 1
Studio review · Thursday · 10:30 AM · 45 minutes. Agenda: Review prototype flow; Confirm first-run wording; Choose pilot feedback measures. Next step: Studio owner shares revised screens by Wednesday.
Based on: Studio review · Draft agendaDelegate for one task
Give an agent a clear job, a bounded set of relevant records, and a scope that does not quietly expand to an entire history.
Separate preparing from acting
Finding a document, drafting a response, sending it, spending money, and making a commitment are different permissions.
Keep an activity trail
See what was requested, which context informed the result, where approval was needed, and what happened next.
MCP is an early integration surface for compatible implementations. Broad agent orchestration and complete delegated policy controls remain future work.
AGENT BUILDERS · NEXT STEP
Design the handoff around the task.
Discuss a system fitUseful details
Questions, answered plainly.
Answers reflect the capabilities and limits described on this page.
Can an agent act without asking me?
The vision distinguishes preparation from consequential actions. Sending, spending, or making a commitment should require human approval when the delegated scope does not explicitly cover it.
Is broad agent orchestration available today?
No. MCP is an early surface for compatible implementations. Broad delegated policies, approvals, and activity trails remain future work.
