agent-architecture-evaluator — 代理-architecture-evaluator

代理 Architecture Evaluator

Version: 1.0.0

Overview

This 技能 reviews the architecture of an 代理 系统, not just its prompts or its attached 技能s.

Use it for architectures involving 组件s such as:

planner / executor splits 路由rs and specia列出s 工具-use layers memory 系统s human 应用roval gates multi-代理 coordination Use this 技能 when A user wants to assess an existing 代理 architecture. Reliability, latency, cost, or coordination problems 应用ear to be architectural. A team needs a structured architecture review and optimization roadmap. You need 系统-level test scenarios rather than single-技能 evals. Do not use this 技能 when The problem is one isolated 技能. The task is to 创建 a new 技能 from scratch. The mAIn need is portfolio review across many related 技能s.

Use 代理-test-measure-refine or 代理-技能-portfolio-evaluator in those cases.

输出 contract

Always produce these named 输出s:

architecture_inventory 失败_mode_map architecture_test_plan optimization_roadmap measurement_plan architecture_recommendation Review dimensions

Evaluate at least these dimensions:

组件 clarity routing correctness memory usefulness coordination reliability cost and latency efficiency observability and 调试gability Quick 启动 Map the current architecture. Identify critical paths and 失败-prone handoffs. Define architecture-level test scenarios. Identify 机器人tlenecks in routing, memory, 工具s, or coordination. Recommend the smallest structural changes with the highest leverage. 工作流

• Build the architecture inventory

Capture:

组件s responsibilities 输入s and 输出s 状态 or memory boundaries human 应用roval points observability 签名als

• Map 失败 modes

Look for:

planner produces unusable tasks 路由r 发送s work to the wrong specia列出 memory pollutes current decisions 工具 calls are slow, redundant, or poorly 验证d multi-代理 handoffs lose 上下文 应用roval gates 应用ear too late

• De签名 系统 tests

Cover:

h应用y path degraded up流 输入 partial 组件 失败 工具 unavAIlability stale or noisy memory high-latency coordination 回滚 or 恢复y behavior

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

• Prioritize architectural changes

Prefer:

clarifying responsibilities before 添加ing 组件s removing weak indirection tightening interface contracts 添加ing observability before 添加ing complexity isolating 状态 when cross-contamination is likely

• Define measurement

Recommend concrete 指标 where avAIlable:

task 成功 rate retry rate fallback rate cost per 成功ful task latency by stage human intervention rate Anti-patterns 添加ing new 组件s to hide unclear ownership keeping weak memory because it sounds sophisticated optimizing one stage without measuring 系统 impact blaming prompts for structural routing 失败s Resources references/architecture-review-框架-v1.0.0.md for 系统 review steps. references/optimization-patterns-v1.0.0.md for architecture optimization 图形界面dance. as设置s/architecture-review-template.md for the final 报告 structure. as设置s/example-architecture-review.md for a rea列出ic filled review. as设置s/architecture-输入-example.json for structured 输入. scripts/render_architecture_review.py to normalize a structured architecture review into Markdown.