Skip to content
Demiton

Agents

Agents are scoped, budgeted tool loops that reason over Civil Memory and deliver a result on their run - either a citation-rich artifact or a proposed change to your organisation’s configuration. They sit alongside Workflows (deterministic verb graphs you configure once and run repeatedly) and Studio (ad-hoc chat): an agent is a bounded, purpose-built reasoning loop with its own budget and its own audit trail.

Find it at /agents in the app.

Two kinds of agent

Every registered agent declares which of two categories it belongs to. The split exists so that agents touching organisation-wide configuration are never trusted with a direct write - a human always reviews the change before it takes effect.

  • Probe agents own one axis of your organisation’s configuration and propose it - they never write it directly. Every side effect lands as a pending approval in your approvals queue, and a probe can re-run itself when it detects drift. Example: a probe that infers which of your systems’ identities map to the same worker.
  • Producer agents consume configuration you have already approved and write reviewable output - never silently finalised, always visible for review before it’s treated as fact. Examples: a measurement agent that verifies drawing quantities against a schedule of rates, or an agent that synthesises a cited tender launch brief.

For the underlying taxonomy - why the split exists and how it prevents unreviewed drift - see Agents under Concepts.

What you’ll find on the surface

  • /agents - the index. Every agent definition available to your organisation, grouped by probe or producer, with a strip of recent run activity.
  • /agents/:slug - an agent’s detail page. Overview (what it does, its budget, what it produces), Runs (its history), and - for probes - Proposals (its pending approvals).
  • /agents/runs/:runId - a single run. Budget burned against budget allowed, a step-by-step timeline that surfaces failures first, and - for measurement agents - the source evidence behind the result.

How it connects to the other surfaces

Agents read Civil Memory the same way Studio and Ask Demiton do, and measurement agents source the underlying documents the same way the platform’s own harvesters do. A probe’s proposals resolve through the same approvals queue as every other proposed change in the system - there is no separate approval mechanism just for agents.

Tiers

Insights and above - the same gate as Memory. Agents require organisation configuration that only exists once your organisation is connected, so there is no Public-tier Agents surface.

Next steps