Software acts on delegated authority.
All capabilities · In build
Automation used to mean a script doing a fixed thing. It now means a system that reads a situation, decides on an approach, and takes actions a person did not enumerate in advance. That shift is a governance problem before it is a technical one, and Unifyhub treats it that way.
Prepare next week’s workshop plan
Read the workshop calendargranted
Draft a proposed schedulegranted
Send the invitationsneeds your approval
Access ends when this task is complete
what the record shows
- asked
- a plan for next week
- allowed by
- workshop coordinator · 14 March
- done
- three sessions drafted
- checked
- no room conflicts
- accepted
- waiting for a reviewer
What an agent is allowed to be
An identity, not an exception
Software acting for a community holds its own identity and appears in the record as itself, never as a shared administrative account that nobody can attribute.
Scoped by delegation
Authority is granted for a stated task, in a stated boundary, for a stated period. Broad standing power is the failure mode, so the system is built to prefer narrow, expiring grants.
Bound by the same policy engine
An agent asks the same question a person's session asks: may this actor take this action on this resource. There is no second, looser path.
Reversible
Revocation is a designed operation with a known blast radius, not a scramble.
How work reaches an agent
Procedural knowledge is packaged
Repeatable know-how is written down as portable procedures with their own tools and checks, so behavior can be reviewed, versioned, distributed, and evaluated instead of living in one prompt somebody wrote once.
Rules constrain the run
Before work starts, the applicable rules are resolved: what this actor may do, what this context requires, what must be validated before completion, and what evidence must exist at the end.
Runtimes are interchangeable
The orchestration layer is independent of any one model provider. Providers are configuration, which is what keeps a community's operating practice from becoming one vendor's roadmap.
Work is scoped and isolated
A delegated change happens in a bounded workspace with a stated objective and a defined way to be accepted or rejected. Nothing lands because the agent believed it was finished.
Evidence is the deliverable
An agentic action is not complete when it produces an output. It is complete when the record shows what was asked, which authority allowed it, what was done, what was checked, and what a reviewer accepted. The record has to show why the system believed the action was permitted, because that is the only part a community can audit after the fact.
The line we hold
Capability that cannot meet the delegation, evidence, and revocation bar stays out of the community's record. That is a real constraint and it will sometimes cost speed. It is the difference between automation a community can govern and automation that governs the community.