This post, written in Chinese by Phodal Huang, announces the desktop release of Routa and explains the thinking behind it. The author says that his work on the ACP protocol and multi-agent collaboration convinced him that a single coding agent is no longer the key problem; what matters is getting several agents to work together inside one engineering system. After a series of multi-agent experiments and a growing body of thinking about harness engineering, he arrived at a formula — “Harness engineering + Coding Agent + Kanban = an AI automated R&D workbench” — and presents Routa as the first fairly complete product embodiment of it.
The post’s recurring question is what “done” means. The author argues that a chat session can carry one act of generation but not a long-lived flow of tasks, in which requirements are split, stages advance, different roles take over, and results are verified, sent back and retried before only some of them pass the gates into delivery.
Key Points
- The author says the hardest problem his team met in moving from agent teams to Kanban was the definition of done (DoD).
- He contrasts the two modes: in a spec-based mode DoD means development is complete (the spec’s
acceptance criteria are met), while in a Kanban mode it means delivery is complete (the card is ready
to enter the
donecolumn). - In his words, “done” in the spec mode is only “dev done” in Kanban, with verification, gates, evidence and state transitions still in between.
- He argues that once DoD means delivery, Kanban becomes a task-level protocol in which moving a card is a state switch that redefines its required inputs, outputs and verification.
- He describes the columns as entry and exit gates rather than a visual queue: each downstream stage checks what the upstream stage delivered and sends the card back if it is incomplete.
- He argues that multi-agent work is not a parallelism problem, since more agents with unclear responsibilities and handoffs only produce noise faster.
- He calls instead for a relay chain of lane specialists, each doing only its own column’s work and explicitly handing responsibility downstream, with gates, rollback and recovery.
- He lists what a team must decide in advance for a codebase AI agents keep rewriting: unbreakable module boundaries, stable contracts, mandatory pre-merge checks, regressions to catch before release, and how progressive disclosure keeps the code understandable to agents.
- He writes that the harness thereby enters the definition of done itself, moving architecture fitness, gate rules, evidence requirements and release conditions forward so the system can decide whether a result moves on or stops.
- He distinguishes a technical “done” (the code has passed every quality gate and can be merged) from a business “done” (the code meets all acceptance criteria and can be released to production).
- He summarises the formula’s division of labour as coding agents supplying execution, Kanban supplying task state, and harness engineering supplying constraints, verification and the definition of done.
Context
The post is a product announcement by the project’s author, framed by his own earlier experiments and writing on multi-agent coding, which it links; its claims about Routa’s design rest on his own account of the project. It applies the author’s thinking on Harness Engineering to task flow and the definition of done.