Multi Agent Orchestration

Multi-Agent Orchestration

Overview

This skill provides a framework for designing and managing multi-agent systems where specialized agents collaborate on complex, multi-stage workflows. It emphasizes clear agent boundaries, structured communication, and robust error isolation.

When To Use

• When tasks are too large or diverse for a single agent (scope creep).
• When specific domain expertise (e.g., database design, UI/UX, security) is required in separate, modular contexts.
• To maintain clean separation of concerns and reduce context window degradation.
• When you need to delegate parallelizable work to maximize throughput.

Topology Patterns

• Hierarchical Supervisor: A central supervisor agent delegates sub-tasks to specialized workers, aggregates their results, and provides final synthesis.
• Swarm Handoffs: Agents pass tasks directly to the next appropriate agent based on completion criteria, forming a chain or graph of expertise.
• Router-Worker: A router analyzes incoming requests and dispatches them to a specific pool of workers based on classification.
• Blackboard: Multiple agents read from and write to a shared persistent state (the "blackboard") until a task objective is satisfied.
• Round-Robin Debate: Agents with opposing viewpoints propose solutions, iterate, and refine based on peer criticism to improve quality.

Communication Protocols

• Agent-to-Agent (A2A): Always use structured message framing.
• Structured Payloads: Encapsulate tasks, constraints, and dependencies in a common JSON format or structured Markdown.
• Return Summaries: Every subagent MUST return a concise summary of work done, resources created, and final status (SUCCESS/FAIL/BLOCKED) before closing the conversation.

Context Isolation & Boundary Hand-offs

• Ephemeral Context: Spawn subagents with only the minimal, high-signal information needed for their specific task.
• Avoid Token Bloat: Do not pass the entire parent conversation history unless strictly necessary. Pass pointers to file locations or artifact links instead.
• Clean State: Each subagent should operate within its own branched workspace to prevent side effects on the parent or other subagents.

Failure Isolation

• Localized Faults: Subagent crashes must be caught by the parent via message timeout or error reporting mechanisms.
• Graceful Retries: Implement retry logic for discrete sub-tasks. If a worker fails, the supervisor should attempt to diagnose the root cause (using Root Cause Tracing) before retrying or pivoting strategy.
• Never Crash Parent: A subagent failure should trigger an alert in the parent agent, not an unhandled exception that propagates to the user.

Task Decomposition Strategies

• Parallel Execution: Use when tasks are independent (e.g., unit tests for different modules, gathering info from multiple docs).
• Sequential Execution: Use when tasks have strict causal dependencies (e.g., design -> implement -> review -> deploy).
• Merge Points: Define clear synchronization points where context from different agents is consolidated, validated, and refined by the supervisor.

State Sharing Patterns

• Shared Artifacts: Write common results to files in the shared artifacts/ directory.
• Blackboard Memory: Use shared databases or documented state files for common configurations or global project context.
• Message Bus: Use the parent agent as the hub for all inter-agent messages.

Anti-patterns

• God Orchestrator: A single agent attempting to do everything; leads to poor specialization and context degradation.
• Circular Dependencies: Agents waiting on each other indefinitely; always define a clear directed acyclic graph (DAG) of task flow.
• Context Explosion: Passing the entire project state to every subagent; use selective scoping instead.
• Silent Failures: Subagents finishing without reporting status; every interaction must have an explicit "done" or "blocked" signal.

Quality Gates

• Pre-Handoff Check: Does the subagent have everything it needs? (Requirements, constraints, deadline).
• Post-Handoff Review: Does the output meet the original task intent? Does it need further refinement before the next step?
• Final Integration: Verify the combined results of all subagents against original user acceptance criteria.

References

• KILO-KIT Core Principles (skills/kilo-kit/SKILL.md)
• Systematic Debugging (skills/systematic-debugging/SKILL.md)
• Architecture Decision Making (skills/architecture/SKILL.md)