A provider-neutral method for supervising multi-agent software engineering with scoped delegation, isolated work, evidence-driven testing, independent review, safe integration, and human-gated publication.
Coding agents make implementation faster, but parallel work introduces new failure modes: overlapping edits, false-green tests, lost context, unsafe environments, and publication without a clear decision owner.
This repository documents the workflow I use to control those risks while keeping agent work practical and auditable. It is a public, sanitized version of a method refined through real software delivery. It contains no private project details.
The primary agent acts as a Supervisor. It audits the real state, divides the objective into bounded outcomes, schedules workers in capacity-aware waves, and validates their evidence. Specialized agents implement, review, test, or integrate. A human retains product and publication authority.
Available agents are capacity, not a target.
| Role | Owns | Cannot approve |
|---|---|---|
| Supervisor | Scope, dispatch, state, and gates | Its own assumptions |
| Development agent | One isolated delivery | Its own integration |
| Independent reviewer | Findings and recommendation | Publication |
| Official integrator | Approved combined state | Unreviewed work |
| Specialized tester | Test plan or evidence | Product release |
| Human validator | Intent, risk, and authority | Nothing delegated blindly |
See roles and responsibilities.
flowchart LR
A[Intake] --> B[Audit]
B --> C[Decompose]
C --> D[Dispatch in waves]
D --> E[Implement and test]
E --> F[Independent review]
F -->|Findings| E
F -->|Approved| G[Safe integration]
G --> H[Integrated validation]
H --> I[Human gate]
I -->|Authorized| J[Publication]
Every arrow is a gate, not an assumption. See the canonical workflow.
The Supervisor maps dependencies, overlapping files, shared environments, risk, and proof requirements before dispatch.
Parallel work is appropriate when outcomes and write areas are independent. Shared files, shared mutable services, or unresolved architecture decisions are serialized. Capacity is reserved for review, integration, and correction.
The complete dispatch method is in task distribution.
A completion message is a claim. Acceptance requires reproducible evidence.
For defects, the preferred loop is:
- reproduce the original behavior;
- prove a meaningful RED;
- apply the smallest coherent correction;
- prove GREEN;
- run focused and proportional regression;
- obtain independent review;
- validate in a safe integrated environment;
- obtain human validation when required.
The method covers unit, integration, API, frontend, browser, negative, visual, and human testing. Read the testing strategy and quality gates.
The core is provider-neutral. Thin adapters translate it to each environment:
- Claude Code adapter uses
CLAUDE.mdand project subagents. - Codex adapter uses
AGENTS.mdand capacity-aware delegation.
In both adapters, the word supervisor has an explicit meaning: stop direct
implementation, load the orchestration contract, audit actual state, and only
then distribute work.
The browser-testing integration references the official Microsoft Playwright Planner, Generator, and Healer agents:
- Planner explores and writes test scenarios.
- Generator converts approved scenarios into executable tests.
- Healer repairs test defects without concealing product regressions.
The definitions are not copied here. This repository documents how the Supervisor scopes their work and reviews their evidence. See the Playwright integration.
A safe interruption preserves work, stops owned services, records external state, and produces a durable handoff. A resumed session audits files, Git, processes, and environments before redispatching unfinished work.
Conversation memory helps; real state decides. See interruption and resumption.
core/ Provider-neutral method
claude/ Claude Code adapter and agent templates
codex/ Codex adapter and AGENTS.md template
templates/ Briefs, journals, test reports, and handoffs
integrations/ Credited external-agent integrations
examples/ Fully fictional workflow examples
- Ask your coding agent to read this repository and summarize the method.
- Select the Claude or Codex adapter.
- Copy and merge the provider template with existing project instructions.
- Customize workspace, data-safety, test, and authority boundaries.
- Run one small task and inspect every gate before expanding parallelism.
- Keep project-specific technical rules in the consuming project.
Start with the Supervisor rules and task brief.
The fictional TeamBoard example shows two implementation waves, RED/GREEN evidence, independent review, interruption, resume audit, browser testing, integration, and human authorization.
External sources, licenses, trademark ownership, and non-affiliation notices are recorded in THIRD_PARTY.md.
The Playwright testing roles come from Microsoft Playwright. Claude Code and Codex behavior is linked only to official provider documentation.
Original content is available under the MIT License. Third-party material remains under its original terms.