Agentic Architecture Review — Development skill for Claude Code

Agentic architecture review

A [Claude Code](https://docs.claude.com/en/docs/claude-code) skill that reviews or redesigns agentic system architectures and implementation plans.

Its bias: **be no more agentic than the problem requires.** Most proposals that arrive labelled "multi-agent" are a fixed pipeline with personas attached. This skill makes you prove otherwise before it agrees, and it checks the parts of an agentic system that architecture documents usually skip: durable state, permissions, tracing, the golden dataset, and evaluations.

What it actually does

Every proposed component gets classified as one of six things:

• Deterministic code or service
• Structured model call
• Code-defined workflow containing model calls
• Tool exposed through a typed contract
• Evaluator or critic node
• Agent with a bounded model-directed tool loop

An agent is only justified when the path cannot be usefully specified in advance and the model must choose a changing sequence of tools to reach an outcome. A sub-agent is only justified by a measured quality gain or a real isolation boundary (identity, permissions, tools, context, model policy, deployment). Personas, pipeline stages, file formats, and deterministic routing do not qualify.

Two principles do most of the work:

• Verification strength caps autonomy. A loop checked by a compiler or a read-after-write confirmation can be given more rope than one checked by the model's own confidence. Self-critique can propose a revision; it should not authorize a consequential action.
• Complexity is earned through error analysis. Establish a baseline, run trials, read the traces, group the actual failure classes, then add the smallest change targeting a demonstrated failure. Not a planner because planners sound thorough.

It also periodically asks the unpopular question: has a model or tool improvement made some of this scaffolding deletable?

Three modes

Mode	Use for
Architecture review	Asse