Agent Orchestration

Selects, sequences, and coordinates cognitive capabilities via OpenClaw
Status: In progress WIP Runtime: OpenClaw Gateway Protocol: OpenAI-compatible HTTP + WebSocket
Overview

Agent Orchestration is implemented by OpenClaw, an open-source AI agent runtime that runs as a daemon on the Mac Mini. OpenClaw provides the Gateway — a control plane that routes messages to agents, invokes skills based on context, and bridges multiple communication channels.

The orchestration model is skill-based: the Guide agent receives user messages, and OpenClaw automatically evaluates registered skills based on their descriptions to determine which capabilities to invoke. The LLM decides when to call a skill based on the conversation context — no explicit routing rules needed.

How Skill Dispatch Works

Each capability in the Workbench is registered as an OpenClaw skill with a SKILL.md file containing YAML frontmatter. The frontmatter includes a name and description that act as trigger phrases. When a user message arrives, OpenClaw evaluates all registered skill descriptions against the context and injects the relevant skill instructions into the agent's prompt.

User: "I have 40 blog posts I need to organize" │ ▼ ┌─── OpenClaw Gateway ───┐ │ │ │ Evaluate skill match: │ │ ✓ librarian-ingestion │ "analyze these documents, │ (score: high) │ profile this corpus..." │ ✗ modeling │ │ ✗ research │ │ │ │ Inject skill into │ │ Guide agent prompt │ └────────┬────────────────┘ │ ▼ Guide responds with Ingestion skill behavior: runs ingest_corpus.py, returns Collection Record
Gateway API

Chat Completions (HTTP)

POST /v1/chat/completions — OpenAI-compatible endpoint. Disabled by default; enable in openclaw.json. Supports streaming. This is how the Vercel frontend connects.

Gateway WebSocket

ws://127.0.0.1:18789 — Full bidirectional control. JSON text frames with req/res/event types. Used by OpenClaw WebChat and native clients.

Tools Invoke (HTTP)

POST /tools/invoke — Direct tool execution without going through the agent. Useful for programmatic access to specific capabilities.

Session Messages (HTTP)

POST /api/sessions/{key}/messages — Send messages to specific sessions. Enables persistent conversations across channels.

Registered Skills

SkillInvocationStatus
Guide (understanding-workbench-guide)Always activeReady to register
Librarian IngestionAuto (model)Ready to register
Librarian EnrichmentAuto (model)Ready to register
Librarian SimilarityAuto (model)Ready to register
Librarian BriefingAuto (model)Ready to register
Architect (PASS)Auto (model)Ready to register
ResearchAuto (model)Ready to register
Configuration

The OpenClaw Gateway configuration lives at ~/.openclaw/openclaw.json on the Mac Mini. Key settings for the Workbench:

Gateway Config (openclaw.json)

SettingValuePurpose
gateway.port18789Default Gateway port
gateway.http.endpoints.chatCompletions.enabledtrueEnable HTTP API for Vercel proxy
gateway.auth.mode"token"Bearer token authentication
gateway.auth.token(from 1Password)Gateway access token
Agency Dial Integration

The Agency Dial — the cross-cutting control that determines how much the system does for vs. with the user — is implemented through OpenClaw's skill invocation controls. Skills can be set to disable-model-invocation: true to require explicit user consent before running (Scaffold Mode), or left as auto-invoked (Instrument Mode). The Guide adjusts its behavior based on the dial setting.

← Back to Dashboard