Prefrontal

Autonomous orchestration with live agent topology, effort routing, anti-goldplating, and CORF trigger.

Paper: J13 — PREFRONTAL: Compounded Intelligence Status: Production (deployed 6+ months) Vanilla OpenClaw: Yes — drop-in installation

What It Does

Acts as the executive control layer for the entire agent system. It tracks every running subagent in a live force-directed topology graph, routes tasks to models by effort tier (cheap for simple, expensive for complex), guards against over-engineered responses (anti-goldplating), fires clarifying questions before expensive work starts (forcing questions), and triggers a structured CORF debate when the agent is stuck in a denial loop. A live call tree is available at GET /api/prefrontal/tree and via the Tinker UI.

Install

1. Copy this folder to ~/.openclaw/workspace/extensions/prefrontal/
2. Add to openclaw.json:

{
  "plugins": {
    "allow": ["prefrontal"],
    "entries": {
      "prefrontal": {
        "enabled": true,
        "config": {
          "autoRoute": true,
          "maxConcurrentWorkers": 8,
          "monitorIntervalMs": 120000,
          "staleThresholdMs": 180000,
          "guardianStaleThresholdMs": 300000,
          "effortRouting": {
            "low": ["ollama/qwen3:latest"],
            "medium": ["anthropic/claude-haiku-4-5"],
            "high": ["anthropic/claude-sonnet-4-6"]
          },
          "featureFlags": {
            "explorationGate": true,
            "antiGoldplating": true,
            "forcingQuestions": true,
            "effortRouting": true,
            "corf": true,
            "faarTracking": true
          }
        }
      }
    }
  }
}

3. Restart gateway

Configuration

Dependencies

• Required: none — all subsystems degrade gracefully when their dependencies are absent
• Optional: All other cognitive plugins benefit from Prefrontal's orchestration. Session store enrichment requires the fork's src/gateway/session-utils.js but falls back silently on vanilla OpenClaw.

How It Works

Prefrontal hooks into subagent lifecycle events (subagent_spawned, subagent_ended), LLM input/output, tool calls, and agent_end to maintain a live TopologyStore graph of all running agents. A background monitor polls for stale/stuck agents and broadcasts markdown status updates to the Tinker UI via ChatEmitter. Six independent subsystems are gated by feature flags: the Exploration Gate prevents redundant work before a task starts; Anti-Goldplating injects a conciseness prompt when it detects over-engineering patterns; Forcing Questions fires clarifying prompts before high-effort work; Effort Router validates that the model tier matches the task complexity; CORF Trigger detects denial-loop patterns and initiates a structured debate; FAAR Tracker logs task completion quality for long-term improvement. The HTTP handler serves the live call tree at GET /api/prefrontal/tree. Topology state is persisted on shutdown and reloaded on restart. Crash recovery state is read and cleared on startup.