A lightweight coordination protocol for parallel AI-assisted development

yimi — Wed, 01 Jul 2026 11:01:29 +0000

AI coding assistants make it much easier to run several pieces of work in parallel. The harder part is coordination: keeping scope, evidence, ownership, and acceptance clear when multiple assistant threads or worktrees are active at the same time.

That is the problem AutoLoop is trying to solve.

AutoLoop is a lightweight, repository-native coordination protocol and evidence-gate toolkit for AI-assisted development. It is built around Markdown templates and PowerShell checks, not a daemon, GUI, autonomous agent runtime, merge queue, or project-management database.

The core loop is intentionally small:

work order -> worker report -> coordinator review / integration review -> closeout

A work order defines what I think of as the Loop Contract: owner, goal, allowed scope, forbidden scope, stop rules, acceptance commands, review gate, commit authority, and final acceptance owner.

A worker report returns the Execution Record: what changed, what was verified, what was not verified, what risks remain, and what should happen next.

Coordinator review or integration review records the Acceptance Decision. That separation matters because "tests passed locally" is not the same thing as live proof, release readiness, or cross-owner integration acceptance.

Why I built it

When you have one assistant working on one branch, chat history may be enough.

When you have several assistant threads working across branches or worktrees, chat-only coordination starts to break down:

a worker may complete local checks but not have authority to accept the work;
one branch can advance the base branch while another branch is still collecting evidence;
a report can describe a state that was true before the final commit, but stale after it;
a task can get split too narrowly, turning integration work into slow process overhead;
project-specific gates can be mistaken for universal workflow requirements.

AutoLoop tries to make these boundaries explicit without taking over the project.

What AutoLoop is

AutoLoop provides a small set of coordination artifacts:

board and coordination-state checks;
work order templates;
worker report templates;
integration review templates;
phase closeout templates;
prompts for coordinator and worker handoff;
PowerShell validation scripts for protocol consistency.

The goal is to keep work reviewable and evidence-based while still letting the project use its own test suite, CI, issue tracker, spec system, review tools, and release process.

What AutoLoop is not

AutoLoop does not automatically dispatch agents, control Codex Desktop threads, merge branches, run deployments, operate hardware, or replace human review.

It is intentionally focused on low-automation coordination:

status and protocol checks;
draftable work orders;
worker reports;
explicit evidence levels;
coordinator acceptance;
integration review when multiple owner lanes are involved.

That boundary is important. The useful part is not "an agent that does everything". The useful part is a clear record of what was allowed, what happened, what was proven, what was deferred, and who can accept the result.

A small example

A coordinator can issue a work order that says:

owner: frontend;
goal: repair one UI flow;
allowed scope: specific files and tests;
forbidden scope: backend API changes, deployment, production data;
acceptance commands: the exact checks the worker must run;
review gate: project-defined;
commit authority: local branch commit allowed;
final acceptance owner: coordinator.

The worker then returns a report with changed files, verification commands and results, unverified items, risks, and the next suggested step.

The coordinator can accept, hold, request a targeted follow-up, or route to integration review. The decision is based on the evidence record, not just a chat summary.

When it helps

AutoLoop is most useful when:

multiple assistant threads or worktrees are active;
tasks need owner lanes and scope boundaries;
local readiness must stay separate from live, production, hardware, or release proof;
report evidence needs to stay synchronized with the actual branch state;
project-specific review gates need to be expressed without hard-coding a particular private tool;
the team wants a lightweight alternative to coordination that lives only in chat.

It is probably overkill for a single tiny change on a single branch.