·research·project·active

belief-graph-orchestrator

Event-sourced multi-scale cognitive control system for autonomous GUI agents.

Problem

A GUI agent that runs longer than a few minutes drifts. Its beliefs about the screen, its intent, and the state of long-running operations all get out of sync. Most agent architectures collapse all of that into a single context window — at which point the cognitive load is on the LLM, not the system.

Solution

An event-sourced belief graph: facts the agent perceives, intentions it forms, and outcomes of actions are all events on a shared log. Belief is the projection. Multi-scale control loops (fast reflex, mid-tactical, slow strategic) consume the same event stream at different granularities.

How

  • Stack: Python.
  • Persistence: event log + materialized belief views.
  • Targets: autonomous GUI agents working in the windows-web / macos-web environments.

Results

Active research. Pairs with windows-web and macos-web-next as the cognitive layer above the controllable-desktop environment.

Lessons

Multi-scale control on a shared log lets you swap the strategic-loop model without retraining the reflex loop. Decoupling tactics from strategy was load-bearing.

Neighborhood

Related

macos-web-nextmacos-web-nextAI systems engineeringAI systems engineeringFull Stack Artificial IntelligenceFull Stack Artificial I...👩🏽‍🌾 The Fertile Cresent👩🏽‍🌾 The Fertile Cre...ComputatrumComputatrumComputatrumComputatrumwindows-webwindows-webbelief-graph-orchestrator