F029: Task Router - Centralized Orchestration Intelligence

Source

MIT Media Lab / Google Research: “Towards a Science of Scaling Agent Systems” (Feb 4, 2026)

Key Research Findings

• Adding more agents often degrades performance
• Centralized orchestration improves parallel task performance by 81%
• Independent multi-agent systems amplify errors by 17.2x vs 4.4x for centralized
• “Tool density” and “decomposability” predict optimal architecture with 87% accuracy

Overview

A task classification and routing service that analyzes incoming tasks and recommends the optimal agent architecture (single agent, centralized swarm, or sequential pipeline). Uses task decomposability and tool density metrics to route decisions.

API

New RPC Method: cstp.classifyTask

{
  "method": "cstp.classifyTask",
  "params": {
    "task": "Review PR #42 and update documentation",
    "tools": ["github", "git", "file_read", "file_write"],
    "constraints": {
      "maxAgents": 3,
      "timeoutSeconds": 300
    }
  }
}

Response

{
  "result": {
    "architecture": "centralized_parallel",
    "confidence": 0.87,
    "reasoning": "Task is decomposable into 2 independent subtasks (code review, doc update) with shared tool dependencies (github, git). Centralized orchestration recommended to contain error amplification.",
    "subtasks": [
      {"id": 1, "description": "Code review", "tools": ["github", "git", "file_read"], "type": "parallel"},
      {"id": 2, "description": "Documentation update", "tools": ["github", "file_write"], "type": "parallel"}
    ],
    "metrics": {
      "toolDensity": 0.67,
      "decomposability": 0.85,
      "predictedErrorRate": "4.4x"
    },
    "alternatives": [
      {"architecture": "single_agent", "confidence": 0.72, "tradeoff": "Simpler but 39% slower for parallel subtasks"},
      {"architecture": "peer_mesh", "confidence": 0.31, "tradeoff": "17.2x error amplification risk"}
    ]
  }
}

Architecture Types

Type	When	Error Rate	Best For
single_agent	Low decomposability, few tools	Baseline	Simple, sequential tasks
centralized_parallel	High decomposability, shared tools	4.4x	PR review + docs, research + writing
sequential_pipeline	Ordered dependencies	2-3x	Build → test → deploy
peer_mesh	⚠️ Rarely recommended	17.2x	Only when tasks are fully independent

Implementation Notes

• Decomposability score: Analyze task description for independent subtask markers (“and”, “then”, “also”)
• Tool density: shared_tools / total_tools across subtasks
• Store classification decisions in CSTP for calibration
• Learn from outcomes: Which architectures actually succeeded?

Integration with OpenClaw

• sessions_spawn could use this to decide single vs multi sub-agent
• Heartbeat could classify queued tasks before execution
• Dashboard: Show architecture recommendations with confidence

Acceptance Criteria

• cstp.classifyTask RPC method implemented
• MCP tool classifyTask exposed
• Tool density and decomposability metrics calculated
• Architecture recommendation with confidence score
• Historical outcome tracking (which recommendations worked)
• Dashboard view showing task classifications