AI Workflow

An AI coding workflow I am researching, and the one Viz Platform is built with. Several AI coding tools each take a role; claiming, implementation, review, and integration all follow task contracts, and work resumes from checkpoints when quota runs out or a process stops.

Illustration, not a product screenshot: several coding agents passing tasks along one delivery flow2026

People set goals and acceptance criteria; assigning tasks, chasing reviews, and resuming interruptions run on their own.

6 roles: roles
1 integrator: single merger

PROJECT EXAMPLES

Screenshots and examples

AI Workflow stages: contract, implementation, review, integration, done — Delivery stages
Redrawn from the project documents: rejected reviews go back to implementation, and only one integrator merges.

AI Workflow delivery flow with supervisors, worker, reviewer, integrator, inspector, and task state — Delivery flow
From the project’s teaching material: task contract → worker → reviewer → integrator, with the inspector flagging exceptions for a person.

01 / Context

The problem

At first I had to open sessions, assign tasks, chase reviews, and pick up interruptions by hand. With several tools editing one repository, write paths and resources had to be separated up front or work would be overwritten. A model saying it was done, or a process exiting with code 0, did not mean acceptance had passed. When quota ran out mid-task, the next session did not know where to continue.

What I built

It now runs on one Mac: the pool scales workers with the claimable tasks, review and integration each have a resident role, and the normal flow completes without the coordinator online. Archived versions 0.1 to 0.6 cover mid-September to mid-October, and none has been formally released. Cross-project and cross-machine management and a general workflow engine are not built.

02 / System

How the system works

01
One task state
Task contracts hold dependencies, write paths, and acceptance criteria. Scheduling state lives in a single JSON file changed atomically under a file lock, and every attempt has its own worktree, lease, and checkpoint.
02
Roles
The coordinator handles priorities and exceptions. Workers claim tasks and submit a fixed commit; reviewers review that commit independently; the integrator merges in a copy, re-runs acceptance, and fast-forwards the main branch; the inspector scans for problems and lists what needs a person.
03
Pool and model routing
Each slot is watched by a launchd-managed supervisor, and the pool scales with claimable tasks. Every task has an ordered list of models; when an account hits its quota or a rate limit, that account is paused and the next candidate takes over.

03 / Decisions

Design choices

Review a fixed commit

A worker submits one fixed commit, the reviewer reviews exactly that commit, and authors cannot write their own approval.

One integrator

However many workers run, only one integrator merges into the main branch, after merging in an isolated copy and re-running acceptance.

Exit code 0 is not done

A model saying it finished, or a clean process exit, does not count. Done requires an acceptance receipt bound to the commit and command times, after integration.

Resume from checkpoints

When quota runs out or a process dies, first confirm the related processes have stopped and keep the workspace, then resume from the checkpoint within a recovery budget. Unclear side effects are left for a person to judge.

04 / Our work

My role

Role：I designed and maintain the workflow. Role contracts, the state tool, the execution pool, and model routing live in the Viz Platform repository; it runs on this one project first before I decide whether to generalize it.

01Designed the role contracts, task state, and delivery flow
02Run it on Viz Platform and change the workflow based on the failures it hits

Multi-agentGit WorktreeTask LeaseCode Reviewlaunchd

Screenshots and examples

Delivery stages

Delivery flow