Agent architecture workbench

What makes an agent want, choose, know, and act?

Animus models an agent as an assembly of replaceable services. It makes the usually implicit control structure—drives, values, power, ideology, cognition, memory, and effect boundaries—inspectable and configurable.

working model · v0descriptive, not yet an orchestratormodules communicate through contracts

Agent cycle

One machine: Observe → Orient → Decide → Act

Outlined sockets are architectural slots; luminous badges are the concrete modules currently installed. OODA describes their recurrent motion through the assembled machine.

Control slot

Harness / control

Owns the agent loop: assembles context, calls inference, selects tools, tracks lifecycle, and coordinates the other modules.

Installed moduleFairyStack with Codex and Claude adapters
OODA participationOwns every phase transition and terminal state
Replacement contractTurn, context, tool, lifecycle, and provider interfaces
MaturityOperational, but control is split across FairyStack and the inner harnesses
Failure pressureHarness-owned model policy or memory weakens interchangeability.

Whole-machine configurations

Values and ideology configure all slots; they are not extra services unless they acquire independent state or ownership.

ReflectiveDeliberate tempo · provenance-first knowledge · reversible effects

OperatorDecisive tempo · budgeted risk · capability-scoped effects

Adversarial criticContradiction search · read-only effects · objections and tests

Applied case: global UI coherence

Header drift reveals page-local orientation without app-global topology.

One primary pageNo global header; sections remain document flow.

Multiple app pagesOne app-owned template renders the shared header everywhere.

Multiple servicesFairyStack owns service discovery, not an injected product shell.

Reusable behaviorHeadless primitives own auth, focus, overflow, and disclosure; the app owns DOM and appearance.

Observe topology → retrieve page-, app-, and platform-scoped knowledge → classify the change as local or shared → mutate the canonical owner → verify every affected route, auth state, and theme.