AI Agents¶
This repository provides two topologies of AI agents:
- Lane Topology — a multi-agent system that classifies each request into a lane and sizes process to match, from a seconds-long mechanical edit to a full plan-build-verify lifecycle. OpenCode-native, with a GitHub Copilot mirror.
- Matrix Topology — the predecessor multi-agent system: a fixed nine-stage lifecycle for production-quality development work. OpenCode-native (canonical), with Claude Code and GitHub Copilot mirrors.
Agent installations are topology-specific. Each topology provides installation instructions for its OpenCode and GitHub Copilot formats.
Installing Agents¶
- Lane Topology — the recommended pattern for new projects, with proportional process sizing.
- Matrix Topology — the structured lifecycle pattern, still published and usable.
Lane Topology¶
The Lane Topology classifies every request into one of five lanes — MECHANICAL, INVESTIGATE, DIRECT, PLAN, or BUILD — by table lookup, and dispatches only the agents that lane needs. It is the successor to the Matrix Topology below, built to close three gaps in that pattern: no middle ground between a trivial change and the full lifecycle, no mode for Socratic planning, and self-reported (not independently executed) test results.
See the Lane Topology page for the full pattern, lane table, agent roster, and installation for both OpenCode (canonical) and GitHub Copilot (mirror).
Matrix Topology¶
The predecessor to the Lane Topology above, and still published and usable: a structured multi-agent pattern for disciplined, lifecycle-driven development. Every request runs the same fixed nine-stage lifecycle — each agent has a defined role, and agents hand off to one another in a prescribed order rather than a single agent doing everything end-to-end.
See the Matrix Topology page for the full pattern description, agent roster, and conductor guide.
Domain Specialists¶
Single-file agents that stand alone, outside any topology. They use the VS Code and GitHub Copilot
.agent.md format, and the three coordinators call the others as subagents. Each is a real file in
agents/, so the file is the authoritative definition.
| Agent | Role |
|---|---|
| api-designer | Designs REST APIs and OpenAPI specifications covering resource modeling, status codes, error shape, pagination, and versioning |
| code-reviewer | Reviews code for correctness, security, maintainability, and test coverage |
| docs-writer | Writes and updates technical documentation such as READMEs, API references, and how-to guides, describing what the code actually does |
| performance-reviewer | Finds performance problems such as inefficient algorithms, N+1 queries, and memory growth, and recommends fixes backed by measurements |
| react-developer | Builds and refactors React components with TypeScript, following the project's existing patterns, accessibility rules, and test conventions |
| security-auditor | Audits code and configuration for security vulnerabilities against the OWASP Top 10 and reports prioritized findings |
| test-engineer | Writes and runs automated tests |
| architecture-coordinator | Designs system architecture by delegating API, security, and performance questions to specialist subagents and synthesizing their answers into one design |
| feature-lead | Leads a feature from design through review by delegating API design, UI work, tests, review, and documentation to specialist subagents |
| review-coordinator | Runs a multi-angle code review by delegating to the code, security, and performance specialists and merging their findings into one prioritized report |
Install one with the GitHub CLI:
A coordinator only delegates to agents it can find, so install the specialists it names alongside it. See Agent Examples for what each one demonstrates, and Subagents for how delegation works.