Multi-Agent Architecture Patterns
Design multi-agent systems with context isolation, explicit handoffs, coordinated execution, and controlled failure handling.
The skill provides architectural guidance and local examples, with no evidence of credential theft, covert exfiltration, or other malicious behavior. However, the filesystem-coordination examples write shared state, create directories, and create lock files without specifying least privilege, sensitive-data handling, user confirmation, isolation boundaries, or rollback, so 9 points are deducted.
The concepts, reference material, and coordination.py are broadly consistent and include failure handling, retries, and circuit-breaker examples. However, key paths lack skill-specific tests; the handoff acknowledgement flow is incomplete; examples rely on simulated execution and placeholders; and some framework snippets lack complete imports or termination behavior. 12 points are deducted; the static cap is 10, so this receives 8.
Activation conditions, topology choices, adjacent-skill boundaries, and non-activation cases are relatively clear, covering supervisor, swarm, and hierarchical systems. Inputs, outputs, non-fit boundaries, and trigger precedence remain insufficiently formalized, and Chinese-language support and mainland-China network fit are not addressed, so 4 points are deducted.
The documentation is well organized and includes progressive topics, examples, gotchas, integration boundaries, references, an MIT license, version, and update date. However, maintenance responsibility is assigned only to a broad contributor group; there is no changelog, explicit update path, dependency-installation guidance, or systematic FAQ, and some material assumes unavailable surrounding context. 5 points are deducted.
The skill can guide topology selection, context isolation, handoffs, consensus, and failure handling, and it includes reusable local coordination code. However, the code contains simulated responses, production replacement points, and framework placeholders; no skill-specific key-path outputs are verified; and substantial engineering work remains, so 9 points are deducted. The static cap is 7, so this receives 6.
The skill includes external research and framework references, and repository-level CI and test materials exist. However, the supplied CI does not cover this skill's key paths, external claims and examples lack verifiable skill-specific test results, and some references support concepts rather than the stated quantitative claims. 6 points are deducted; the static cap is 5, so this receives 4.
- Do not treat scripts/coordination.py as production-ready: it contains synchronous simulation, placeholder logic, and incomplete handoff acknowledgement.
- Before using filesystem coordination, add path restrictions, a permission model, concurrency guarantees, sensitive-data rules, and recovery procedures.
- Verify the “approximately 15x cost,” LangGraph performance differences, and other quantitative claims against the original research; this review is static only.
- Before adopting a framework example, complete dependencies, imports, graph termination conditions, and framework-version compatibility.
What does this skill do, and when should you use it?
This skill guides developers designing multi-agent systems around supervisor, peer-to-peer swarm, and hierarchical architectures. Its central principle is context isolation: each agent should receive the context needed for its subtask without carrying unnecessary accumulated history. It covers parallelization, handoffs, consensus, validation, and failure mitigation. It is an architecture guide rather than an executable multi-agent runtime.
Explains how to select supervisor, swarm, or hierarchical topologies based on decomposition and coordination needs; designs explicit handoff and state-passing protocols; compares full-context delegation, instruction passing, and filesystem memory; recommends weighted voting, debate protocols, validation checkpoints, retries, circuit breakers, and time-to-live limits; and provides Python pseudocode, architecture diagrams, and references for LangGraph, AutoGen, and CrewAI.
- An agent-platform engineer deciding whether a complex task benefits from multiple isolated contexts.
- A research-team developer coordinating parallel source searches, document analysis, and synthesis.
- A customer-service developer routing billing, technical, sales, and general requests through explicit handoffs.
- A production team diagnosing supervisor saturation, telephone-game information loss, or cascading agent errors.
- An engineering team structuring a large workflow across strategy, planning, and execution layers.
What are this skill's strengths and limitations?
- Covers supervisor, peer-to-peer swarm, and hierarchical architectures.
- Treats context isolation as the primary design benefit of multi-agent systems.
- Provides concrete guidance for handoffs, consensus, validation, and failure handling.
- Addresses supervisor bottlenecks, information degradation, coordination overhead, and error propagation.
- Uses platform-agnostic principles with transferable Python pseudocode examples.
- It is primarily architecture guidance and does not provide a complete executable framework implementation or test suite.
- Multi-agent systems add token cost, communication overhead, and latency; the skill's gotchas recommend budgeting for roughly 15x baseline cost.
- Framework-specific implementation requires consulting LangGraph, AutoGen, or CrewAI documentation.
- The source does not provide project-specific performance results or deployment configurations.
How do you install this skill?
Copy the complete skills/multi-agent-patterns directory into a compatible skills directory, for example: mkdir -p .codex/skills && cp -R skills/multi-agent-patterns .codex/skills/. It can also be copied into .claude/skills/ or the generic .agents/skills/ directory. Do not flatten SKILL.md into a standalone file. The source does not specify additional dependency installation.
How do you use this skill?
Use it in an Agent Skills-compatible host with a request such as: Choose a multi-agent architecture for a system that performs parallel research, fact-checking, and report writing, then define its handoff and validation protocols. It is intended to activate for context isolation, supervisor or swarm coordination, explicit handoffs, parallel execution, or deciding whether multiple agents are justified.
How does this skill compare with similar options?
The skill compares supervisor/orchestrator, peer-to-peer/swarm, and hierarchical architectures: supervisors fit centralized control and human oversight, swarms fit flexible exploration, and hierarchical systems fit large projects with strategy, planning, and execution layers. It also compares full-context delegation, instruction passing, and filesystem memory as isolation mechanisms.