Routa is an open-source multi-agent orchestration platform, published as the routa-js repository,
that 2026 年，万物皆 Coding Agent 的平台工程（A2A / ACP / MCP / Skill） presents as a practical
reference for orchestrating coding agents. It uses a hybrid of the two orchestration modes that post
describes: an external orchestrator that drives coding agents as separate ACP processes, and a
Workspace Agent that can act as a macro-coordinator over sub-agents as an optional mode.
Its author, Phodal Huang, later announced a desktop version in Routa 桌面版发布：内建 Harness 工程的 AI Coding 研发协作工作台, describing Routa there as an AI coding collaboration workbench with built-in harness engineering and as the first fairly complete product form of his formula “Harness engineering + Coding Agent + Kanban = an AI automated R&D workbench”. That post says Routa aims to reconnect the flows of tasks, responsibilities and evidence, rather than to be a chat-style coding agent or only a container for multiple agents.
Capabilities
The architecture the platform-engineering post describes has five layers. A protocol gateway accepts Agent2Agent Protocol (A2A) for federation, Model Context Protocol for tools, the Agent Client Protocol (ACP) for processes and REST for management. Below it, an orchestration engine layer holds an Orchestrator that decomposes tasks, manages dependencies and schedules agents; an EventBus that carries event-driven communication and handles agents’ reports to their parent; and a Skill Matcher that matches agents to tasks by the task’s characteristics. A state layer keeps tasks, agents, notes and traces, an ACP process manager spawns, prompts, cancels and subscribes to agents, and the bottom layer is the coding agents themselves — the post names Claude Code, OpenCode, Gemini and Codex.
Work is divided among three roles with explicit boundaries: a Coordinator that plans, decomposes and summarises but does not write code directly, an Implementor that builds each task without widening its scope, and a Verifier that accepts and reviews work but does not modify code. That post gives Claude (marked “smart”) as the typical provider for the Coordinator and Verifier, and OpenCode (marked “fast”) for the Implementor.
The desktop-release post centres on Routa Kanban, which it describes as at once a task board, the
interface for assigning work to agents, and a task state machine with built-in quality gates, stage
constraints and domain specialists. Its columns — backlog, todo, dev, review, done and
blocked — work as entry and exit gates: a card entering backlog must carry machine-readable YAML
with a problem statement, acceptance criteria, scope, dependencies and an INVEST check; todo asks for
an execution plan, key files, a dependency plan and risks; dev asks whether the work is ready to code
and requires development deliverables after implementation; review checks that evidence is complete,
verification independent and scope contained; and done requires an approved review result. A
downstream column can send an incomplete card back upstream.
Cards are moved between columns by lane-specialist agents, each doing only its own column’s work,
verifying the upstream output first and handing off explicitly with move_card. The post lists a
KanbanTask Agent that splits natural-language requirements into backlog-ready cards without
implementing them; a Backlog Refiner that turns rough cards into executable stories; a Todo
Orchestrator that checks backlog output and adds the execution plan; a Dev Crafter that implements,
commits and attaches development evidence; a QA Frontend specialist that adds visual-regression and
screenshot evidence for UI work; a Review Guard that acts as the final quality gate, independently
checking acceptance criteria, tests, Git state, scope and the Entrix budget; a Blocked Resolver that
reroutes blocked work; and a Done Reporter that writes a completion summary. A PR Publisher prepares
branches and opens pull or merge requests as a delivery-side specialist outside the basic board flow,
and a general Kanban Workflow specialist takes over only on columns that have no lane specialist.
Around Routa the author grew several further components: Entrix, which turns quality rules, architecture constraints and verification steps into executable guards; a Harness Monitor that keeps watching quality and change while several agents work on one codebase; a Harness Dashboard that visualises those signals; and Routa Kanban itself.
Adoption & Ecosystem
The desktop-release post reports that Routa has close to 500,000 lines of code, almost all of it generated by AI.
The platform-engineering post lists the practices the design rests on: external systems trigger tasks through webhooks or A2A; providers are pluggable, so the same role can be filled by different coding agents and switched at runtime; all state changes pass through the EventBus to support asynchronous collaboration; and role instructions and behavioural boundaries are defined as Skills written in Markdown with YAML frontmatter.
Harness Engineering 实践指南：落地探索的三大原则 uses Routa.js as a case study in Harness Engineering. It describes the project as both the working environment for AI agents and a codebase that depends on those agents taking part in its own development, with several agents — it names Kiro, GitHub Copilot, Augment and Claude Code — collaborating in the same repository under shared engineering rules and quality gates. That post sorts the project’s practices under three headings:
- System legibility: an AGENTS.md file sets coding standards, testing strategy, Git discipline and the co-author format for AI commits, and Specialist configurations define each agent’s role and boundaries so that roles such as Developer and Verifier follow the process.
- Engineering defences: two layers of Git hooks, a fast pre-commit check and a pre-push check that returns structured error feedback, and a lint policy that lets some rules drop to warnings so that multi-agent collaboration is not blocked while the problems are still recorded for follow-up.
- Automated feedback loops: an Issue Enricher prepares context and proposes solutions for new issues; a Copilot Complete Handler moves draft pull requests to ready-for-review and triggers a review by an Augment agent; an Issue Garbage Collector periodically clears stale issues, with local issue files providing persistent context; and a Kanban system triggers agent tasks and requires verifiable artifacts such as test results or code diffs.