How Multi‑Agent AI Redefines Development: The Harness‑Powered Routa Workbench

Background

Early experiments showed that moving a single Coding Agent to a CLI was insufficient; the real challenge is coordinating multiple agents within a unified engineering system. The solution is to combine a Harness engineering framework, Coding Agents, and a Kanban‑based workflow to create an AI‑automated development workbench.

Definition of Done (DoD) in Multi‑Agent Workflows

DoD differs between two modes:

Spec‑based mode : "Done" means the implementation satisfies all acceptance criteria defined in the specification.

Kanban‑based mode : "Done" means the card has reached the done column, indicating delivery completion.

Thus, development completion (Spec‑mode) becomes a prerequisite for delivery completion (Kanban‑mode). Between these stages the system enforces verification, gatekeeping, evidence collection, and state transitions.

Routa Harness Engineering Stack

The stack provides reusable safeguards for multi‑agent projects:

Entrix – translates quality rules, architectural constraints, and verification steps into executable safeguards. Repository: https://github.com/phodal/entrix Harness Monitor – continuously observes quality and changes when multiple agents modify the same codebase.

Harness Dashboard – visualizes signals from the monitor and other components.

Routa Kanban – integrates requirements, status, handoffs, and gatekeeping into a single task flow.

The codebase has grown to nearly 500 000 lines, almost entirely generated by AI.

Kanban as a Task Protocol

Kanban columns become explicit engineering semantics and boundaries for specialist collaboration. Each column enforces entry and exit gate conditions:

Backlog : requires a machine‑readable YAML containing problem statement, acceptance criteria, impact scope, dependencies, and INVEST checks.

Todo : expands into an execution plan, key files, dependency plan, and risk description.

Dev : confirms readiness for coding; after implementation, all development artifacts must be submitted.

Review : validates that evidence is complete, verification is independent, and scope is controlled.

Done : requires an APPROVED review and acceptance by the delivery system.

Movement between columns is not visual only; it triggers validation and may reject a card to a previous column if gate conditions are not met.

Lane Specialists for Multi‑Agent Collaboration

Each Kanban lane is owned by a specialist that validates upstream output before performing its own tasks and then explicitly hands off responsibility using move_card:

Backlog Refiner : refines rough cards into executable stories, completing canonical YAML, acceptance criteria, dependencies, and INVEST checks.

Todo Orchestrator : validates backlog output and adds execution plans, key files, and risk assessments.

Dev Crafter : implements code, submits changes, and records development evidence before sending to review.

QA Frontend : adds UI‑specific verification evidence for visual regression.

Review Guard : final quality gate that independently verifies acceptance criteria, tests, Git status, scope control, and Entrix budget.

Blocked Resolver : handles blockers, rerouting unclear requirements or environment issues to appropriate stages.

Done Reporter and PR Publisher : finalize delivery summaries and create pull/merge requests.

Redefining "Done" in an AI‑Driven Project

For long‑lived projects, "Done" has two dimensions:

Technical Done : code passes all quality gates and can be merged into the main branch.

Business Done : code meets all acceptance criteria and can be released to production.

In continuously AI‑rewritten codebases, the system must enforce:

Architecture tests that prevent boundary violations.

Contract stability for interfaces and data structures.

Mandatory quality gates before merging.

Regression risk identification prior to release.

Agent understandability so agents can read, locate, and modify code incrementally.

These concerns are embedded in the Harness framework as fitness checks, gate rules, evidence requirements, and release conditions, allowing the system to decide whether to proceed or halt.

Conclusion

Routa integrates three pillars—Harness engineering, Coding Agents, and Kanban—to form an AI‑automated R&D workbench. Coding Agents provide execution capability, Kanban supplies task state, and Harness supplies constraints, verification, and a rigorous Definition of Done. Only when all three are tightly coupled does AI coding move beyond rapid code generation to reliable, collaborative, and deliverable software development.