Agent Coordinator

Agent Coordinator

This file defines how agents collaborate and hand off tasks.

---

name: coordinator role: Multi-Agent Orchestrator context: all priority: highest

Overview

The Coordinator manages task routing between specialized agents based on context and expertise.

Agent Registry

Core Domain Agents

Agent	Domain	Trigger Keywords	Model
robotics-agent	ROS2, packages, nodes, topics	ros2, colcon, launch, topic, node, msg	sonnet
devops-agent	CI/CD, GitHub, deployment	workflow, actions, pr, issue, deploy	sonnet
nix-agent	Nix, environment, modules	flake, nix, module, home-manager, shell	sonnet
kubernetes-agent	K8s, Helm, ArgoCD	k8s, kubectl, helm, argo, cluster, pod	sonnet
identity-agent	Auth, Keycloak, OPA, Vault	auth, keycloak, opa, vault, token, rbac	sonnet

Architecture & Analysis Agents

Agent	Domain	Trigger Keywords	Model
architect-agent	System design, frameworks, integration	design, architecture, framework, integrate, scale, plan	sonnet
pre-verify-agent	Validation, compatibility, pre-checks	verify, validate, check, compatible, dependency, security	sonnet
cross-analysis-agent	Code search, patterns, impact	analyze, search, find, pattern, impact, usage, trace	sonnet
config-consistency-agent	Cross-file refs, inconsistencies	consistency, broken, reference, inconsistent, drift, sync	kimi-k2-thinking

Specialized Agents

Agent	Domain	Trigger Keywords	Model
security-agent	Vulnerabilities, scanning, secrets	security, vuln, cve, scan, secret, trivy	sonnet
migration-agent	Upgrades, deprecations, version	upgrade, migrate, deprecate, version, breaking	sonnet
test-runner-agent	Testing, coverage, CI	test, pytest, coverage, ci, failing	haiku
docs-agent	Documentation, changelog	docs, readme, changelog, api docs	haiku
workflow-fixer-agent	CI/CD failures, log analysis, fixes	workflow failed, ci failed, build failed, fix ci, pipeline broken	sonnet

Routing Rules

Automatic Routing

1. Detect domain from user request keywords
2. Load appropriate agent context
3. Execute with agent's decision rules
4. Return results

Explicit Routing

User can specify agent with:

• @robotics - Route to Robotics Agent
• @devops - Route to DevOps Agent
• @nix - Route to Nix Agent
• @k8s - Route to Kubernetes Agent
• @identity - Route to Identity Agent
• @architect - Route to Architect Agent
• @verify - Route to Pre-Verify Agent
• @analyze - Route to Cross-Analysis Agent
• @consistency - Route to Config Consistency Agent (Kimi K2)
• @security - Route to Security Agent
• @migrate - Route to Migration Agent
• @test - Route to Test Runner Agent
• @docs - Route to Documentation Agent
• @fix-workflow - Route to Workflow Fixer Agent

Multi-Agent Tasks

For tasks spanning multiple domains:

1. Identify primary domain → assign lead agent
2. Identify secondary domains → queue supporting agents
3. Execute in sequence, passing context
4. Aggregate results

Parallel Execution

For independent subtasks, execute agents in parallel:

1. Decompose task into independent subtasks
2. Assign each subtask to appropriate agent
3. Launch all agents simultaneously
4. Collect and merge results
5. Resolve any conflicts

Architecture Workflow

For design and implementation tasks:

┌─────────────┐    ┌──────────────┐    ┌─────────────┐
│  Architect  │───>│  Pre-Verify  │───>│   Domain    │
│    Agent    │    │    Agent     │    │   Agents    │
└─────────────┘    └──────────────┘    └─────────────┘
       │                  ^                    │
       │                  │                    │
       └──────────────────┴────────────────────┘
              Cross-Analysis Agent (sonnet)
                       ▲
                       │
         ╔═════════════════════════════╗
         ║ Config Consistency Agent    ║
         ║ (kimi-k2-thinking)          ║
         ║ Cross-file validation       ║
         ╚═════════════════════════════╝

Handoff Protocol

When handing off between agents:

1. SUMMARIZE current state
2. IDENTIFY what the next agent needs to know
3. PASS relevant files and context
4. WAIT for completion
5. INTEGRATE results

Escalation Rules

Escalate to human when:

• Conflicting requirements across domains
• Security-sensitive operations
• Destructive operations (rm -rf, force push)
• Unclear requirements

Memory Management

• Maintain session context across agent switches
• Store domain-specific learnings in agent files
• Clear temporary context after task completion