Enterprise Agents: A modular AGENTS.md governance kit for multi-agent coding

Michail Peterlis — Thu, 12 Nov 2026 10:21:39 +0000

For the past few weeks I've been using several agents privately and at the company I'm working for. In that time, across several harnesses, I generated quite a few chats used to plan, architect, implement, release, and deploy various tools and modules. I assembled those conversations and outcomes into a set of AGENTS.md files which I wanted to share. It is still a work in progress and will be extended in the future. It is MIT-licensed and free to use. Any feedback is welcome.

michail-peterlis / AGENTS.md

A modular, enterprise-grade AGENTS.md framework for reliable multi-agent software engineering with explicit roles, policies, workflows, review gates, and progressive context loading.

Enterprise Agents

A modular AGENTS.md framework for disciplined multi-agent software engineering.

Coding agents can produce code quickly. Reliable software engineering also requires:

explicit edit ownership
scoped acceptance criteria
independent review
evidence preservation
test integrity
architecture discipline
production safety
progressive context loading

Enterprise Agents is an opinionated governance framework built around the open AGENTS.md convention. It is not a new AGENTS.md standard.

What this repository provides

AGENTS.md       universal repository rules
.agents/roles/ who the agent is
.agents/policies/
                cross-cutting engineering rules
.agents/skills/
                just-in-time workflows
nested AGENTS.md
                component-specific rules
task/plan       what is being done now

The root AGENTS.md acts as a router for the framework unless a particular harness provides native support for additional files.

Compatibility statement

This project uses the open AGENTS.md convention for repository-level instructions. The .agents/roles, .agents/policies, .agents/skills, templates, and related structure are an opinionated extension provided by this framework and are not part of the core AGENTS.md…

View on GitHub

How it is structured

Instead of packing every rule into one huge AGENTS.md file that consumes the context window, the root AGENTS.md acts as a lightweight router for progressive context loading across .agents/:

AGENTS.md          universal repository rules
.agents/roles/     who the agent is (architect, implementer, challenger, etc.)
.agents/policies/  cross-cutting engineering rules
.agents/skills/    just-in-time workflows and POSIX sh coordination helpers
.agents/templates/ reusable task, state, handoff, and review records

Key principles

Progressive context loading: Agents read only the root AGENTS.md, their assigned role, and task-relevant policies or skills rather than preloading every document.
Independent review: Separates implementation from adversarial review (challenger role) with explicit single-writer ownership and acceptance tiers (T0–T2).
Harness-independent: Works across single sessions, delegated subagents, headless staged runs, or live peer sessions coordinating over a git-ignored .agents/workspace// file channel.
Dual-edition releases: Ships both the full kit and a condensed lite edition for smaller context windows, for local models.

DEV Community: Michail Peterlis