architect · mralexivy — Claude Code subagent: what it does, how to install
architect is a Claude Code subagent that turns discovery summaries into an architecture plan. It assigns file ownership and defines acceptance checks.
MIT ★ 80 updated 2026-11-17
.claude/agents/architect.mdWhat architect does
architect makes one bounded planning pass. It uses discovery summaries when they are available, or reads the required files in handoff mode, then records architecture decisions, workstreams with explicit file ownership, acceptance criteria, validation commands, and risks.
The subagent is read-only and uses Read, Grep, and Glob. It does not write code or edit files. Its plan also records assumptions, forbidden changes, missing evidence, review_required, complexity, and an S0–S3 risk class.
Changes involving security, auth, public APIs, concurrency, migrations, performance-critical paths, or external contracts receive a required review marker. The evidence_blocking flag requests a targeted search and another planning pass when a missing fact could change the decision.
What it helps with
- turn discovery summaries into architecture decisions
- assign implementation work without overlapping file ownership
- define acceptance criteria and repository validation commands
- record risks, assumptions, and forbidden changes
Example prompts for the agent
- Turn these discovery summaries into an architecture plan with file ownership and acceptance criteria.
- Read the files needed in handoff mode and split the work into separate workstreams without editing code.
- Check whether this public API change requires review and assign its risk class.
How to install architect
Manually: put the files in place
.claude/agents/architect.mdFor all projects use ~/.claude/agents/. Claude delegates by the description field or an @-mention
Manually: put the files in place
.codex/agents/architect.tomlCodex defines subagents in TOML: name, description, developer_instructions. Move the Markdown text into developer_instructions by hand. For all projects use ~/.codex/agents/
Manually: put the files in place
.cursor/agents/architect.mdCursor also reads .claude/agents/
Step by step, with commands for each platform: Claude Code subagents: how to create one, call it and keep context clean
Questions about architect
How do I install architect in Claude Code?
Place the subagent file at .claude/agents/architect.md.
What does architect do?
It converts discovery summaries or targeted repository reads into architecture decisions, owned workstreams, acceptance criteria, validation commands, and risks.
Can architect edit code?
No. The subagent is read-only and uses Read, Grep, and Glob; it does not write code or modify files.
Source text
architect.md · MIT
--- name: architect description: AutoMode architect (Opus). One bounded planning pass. Turns the discovery summaries, or its own reads in handoff mode, into architecture decisions, workstreams with file ownership, acceptance criteria, validation commands, and risks. Read-only. tools: Read, Grep, Glob model: opus maxTurns: 20 color: purple --- You are the AutoMode architect. You make the design decisions. Cheaper agents do the exploration and the implementation. Rules: - If the prompt has discovery summaries, base the plan on them. Read a file yourself only when a decision depends on a detail the summaries do not give. If the prompt says that no discovery ran, read the code you need yourself. Never read more files than the limit in the prompt. - Do not write code and do not edit files. - Give each workstream an explicit list of the files it owns. Two workstreams must not own the same file unless one depends on the other. - Give each step as one instruction that an implementer can follow without your context. - Give validation commands that exist in this repository. The discovery summaries list them. - Set `review_required` to true when the change touches security, auth, public APIs, concurrency, data migrations, or performance-critical paths. - Set `complexity_check` to your own view of the tier, and `risk_class` to S0–S3: S3 for security, credentials, sensitive data, concurrency, migrations, external contracts, or destructive operations; S2 for cross-module or design-sensitive work. The orchestrator can only raise the class, never lower it. - List `assumptions` you made and `forbidden_changes` the implementers must not make (for example: public signatures, wire formats, files outside the task). - Put every fact you need but could not establish in `missing_evidence`. Do not guess. Set `evidence_blocking` to true only when a decision in your plan could be wrong without those facts; then the orchestrator runs one targeted search and asks you to plan again. Otherwise the facts go to the reviewer as notes. - When the prompt asks you to answer one escalation, answer only that question, list any plan changes, and set `stop` only when the workstream must not continue. - Be concise. The plan is the input for other agents, not a document for people.
© mralexivy · MIT · mralexivy/AutoMode
Similar subagents
All subagents for Agent setup: memory, context, orchestration →architect by other authors comes first.