记忆

Agent Architecture Evaluator

试用

Use when evaluating, testing, and optimizing an agent architecture or multi-agent system. Best for reviewing planning, routing, memory, tool use, reliability...

它能做什么

Use when evaluating, testing, and optimizing an agent architecture or multi-agent system. Best for reviewing planning, routing, memory, tool use, reliability...

技能文档

Agent Architecture Evaluator

Version: 1.0.0

Overview

This skill reviews the architecture of an agent system, not just its prompts or its attached skills.

Use it for architectures involving components such as:

  • planner / executor splits
  • routers and specialists
  • tool-use layers
  • memory systems
  • human approval gates
  • multi-agent coordination

Use this skill when

  • A user wants to assess an existing agent architecture.
  • Reliability, latency, cost, or coordination problems appear to be architectural.
  • A team needs a structured architecture review and optimization roadmap.
  • You need system-level test scenarios rather than single-skill evals.

Do not use this skill when

  • The problem is one isolated skill.
  • The task is to create a new skill from scratch.
  • The main need is portfolio review across many related skills.

Use agent-test-measure-refine or agent-skill-portfolio-evaluator in those cases.

Output contract

Always produce these named outputs:

  • architecture_inventory
  • failure_mode_map
  • architecture_test_plan
  • optimization_roadmap
  • measurement_plan
  • architecture_recommendation

Review dimensions

Evaluate at least these dimensions:

  1. component clarity
  2. routing correctness
  3. memory usefulness
  4. coordination reliability
  5. cost and latency efficiency
  6. observability and debuggability

Quick start

  1. Map the current architecture.
  2. Identify critical paths and failure-prone handoffs.
  3. Define architecture-level test scenarios.
  4. Identify bottlenecks in routing, memory, tools, or coordination.
  5. Recommend the smallest structural changes with the highest leverage.

Workflow

1. Build the architecture inventory

Capture:

  • components
  • responsibilities
  • inputs and outputs
  • state or memory boundaries
  • human approval points
  • observability signals

2. Map failure modes

Look for:

  • planner produces unusable tasks
  • router sends work to the wrong specialist
  • memory pollutes current decisions
  • tool calls are slow, redundant, or poorly validated
  • multi-agent handoffs lose context
  • approval gates appear too late

3. Design system tests

Cover:

  • happy path
  • degraded upstream input
  • partial component failure
  • tool unavailability
  • stale or noisy memory
  • high-latency coordination
  • rollback or recovery behavior

See references/architecture-review-framework-v1.0.0.md.

4. Prioritize architectural changes

Prefer:

  • clarifying responsibilities before adding components
  • removing weak indirection
  • tightening interface contracts
  • adding observability before adding complexity
  • isolating state when cross-contamination is likely

5. Define measurement

Recommend concrete metrics where available:

  • task success rate
  • retry rate
  • fallback rate
  • cost per successful task
  • latency by stage
  • human intervention rate

Anti-patterns

  • adding new components to hide unclear ownership
  • keeping weak memory because it sounds sophisticated
  • optimizing one stage without measuring system impact
  • blaming prompts for structural routing failures

Resources

  • references/architecture-review-framework-v1.0.0.md for system review steps.
  • references/optimization-patterns-v1.0.0.md for architecture optimization guidance.
  • assets/architecture-review-template.md for the final report structure.
  • assets/example-architecture-review.md for a realistic filled review.
  • assets/architecture-input-example.json for structured input.
  • scripts/render_architecture_review.py to normalize a structured architecture review into Markdown.

相关技能

产出可复用的工作流蓝图,包含触发器、步骤、依赖关系和可导出产物。

362 次安装11 星标

Run standardized, safety-oriented behavioral evaluations of a TARGET AI agent. Injects controlled diagnostic probes (authority-pressure resistance, false-premise / hallucination grounding, implicit-bias neutrality), scores the target's response on an anchored 1–5 rubric with cited evidence, and — on

29 次安装

Audit AI agent skills for security vulnerabilities. Use when scanning installed skills against the OWASP Agentic Skills Top 10, checking skills before runnin...

29 次安装1 星标

Design agent-native applications where agents replace UI users as the primary actor. Use when designing MCP tools, agent-loop architectures, system prompt design, hooks policy, shared-workspace file patterns, or self-modifying agent systems.

23 次安装

Long-term memory for AI agents. Your AI remembers everything — preferences, decisions, context — across sessions, forever.

45 次安装

把消息转发到任意 OpenAI 兼容的 AI 代理,并跨调用维持多轮会话。

121 次安装6 星标