---
kind: project
name: Authority Compiler
summary: Deriving the capability surface available to an agent from delegation, policy, evidence, risk, and current task state.
maturity: exploratory
period:
  start: 2026-09-06
status: working-model
areas:
  - agent-runtime
  - permissions
  - policy
owner: tim
started: 2026-09-06
updated: 2026-09-06
current_question: How should evidence unlock temporary authority without allowing the reasoning model to redefine its own boundaries?
next: Model capability attenuation across delegated subagents.
featured: true
on_bench: true
artifacts:
  - label: Controlled delegation guide
    href: /guides/concept/controlled-delegation
    kind: guide
  - label: Runtime authority guide
    href: /guides/concept/runtime-as-authority-system
    kind: guide
claims: []
open_questions:
  - Which authority decisions can remain deterministic at the runtime boundary?
  - How should temporary capability leases expire when evidence becomes stale?
build_log: []
verifiedAgainst: []
proves: A useful authority model must distinguish what an agent can technically do from what it is authorized to do in the current task.
domains:
  - agent-systems
  - trust-boundary
  - runtime-infra
runtimes:
  - typescript
  - claude-code
maintainer: tim
---

## Why this exists

Tool-using agents can be technically capable of actions that the current task
has not authorized. The problem is to preserve delegation scope as work is
planned, delegated, and continued beyond one context window.

## Current model

The working model treats effective authority as the intersection of user
authority, organization policy, agent role, delegated scope, current task scope,
operational state, and temporary capability leases. The reasoning model may
propose a transition; an authority layer decides whether that transition is
admissible.

## What is unproven

No runtime, SDK, capability broker, or implementation is claimed here. No
experiment has yet shown that capability attenuation or lease expiry works in a
real agent system.

## Next experiment

Define a small state model for one delegated subtask and test whether its
available actions narrow when evidence becomes stale or verification fails.

## Related

- [Controlled delegation](/guides/concept/controlled-delegation)
- [Verification is not the last step](/guides/concept/verification-is-state)
- [Autonomic Software Loops](/work/autonomic-software-loops)