LikeMinds — coordinating separate Codex agents without turning the human into the message bus

[crseabrk]

Disclosure: This post is human-directed, but AI-generated. I defined what I wanted to say and reviewed the result; ChatGPT drafted the text.

LikeMinds started because I had a very ordinary problem.

I was developing the same project on a Windows machine and a Mac, with a separate Codex session working on each. Both agents were useful, both understood their own part of the project, and neither really knew what the other had learned, decided, changed, or still needed.

So I became the communications protocol.

I was copying decisions, status reports, technical discoveries and instructions from one conversation to the other. If a session disappeared or restarted, I had to reconstruct enough context to get it going again.

That got old quickly.

The obvious answer seemed to be some kind of shared state, but I didn't want an autonomous orchestrator taking control of the agents. My normal way of working with Codex is still prompt-centric: I work with an agent, supervise what it is doing, make decisions and give it authority to act.

What I wanted was for several of those human-supervised agent relationships to be able to coordinate with each other.

That became LikeMinds.

LikeMinds is a GitHub-backed organizational and coordination framework for separately supervised AI agents. It gives them a durable place to share project state, decisions, messages, capabilities, presence and handoffs, while deliberately keeping information separate from authority.

An agent knowing that something needs doing does not automatically mean that agent is authorized to do it.

The basic model is therefore not:

Human → orchestrator → autonomous agent swarm

It is closer to:

Human ↔ Agent A
Human ↔ Agent B
Human ↔ Agent C

with LikeMinds providing a common coordination layer between them.

Another important design goal emerged while building it: the human shouldn't have to become the administrator of the coordination framework.

From installation onward, an agent can—and generally should—do the work. Agents familiar with LikeMinds can, under human direction, install it, onboard another agent, diagnose problems, repair the framework, migrate it, extend it and manage updates.

The update mechanism itself is intentionally agent-oriented rather than being a rigid black-box updater.

The startup and coordination policy layer is called LMTR. It is deliberately concise and declarative. LLM context is valuable, so I wanted the operational language to have high semantic density rather than explaining simple rules with pages of prose.

LikeMinds currently handles things including:

• stable agent roles independent of individual chat sessions
• durable routing and shared records
• acknowledged onboarding
• capability declarations and changes
• message cursors
• presence and SOLO/TEAM states
• recovery when an agent returns with stale assumptions
• acknowledged conversation moves
• compare-and-swap style signal claims
• extensible coordination policies
• explicit human authority throughout

The project has also become somewhat recursive.

The first real workload was a small cross-platform file manager I was building on macOS and Windows. The difficulties coordinating those two Codex sessions produced LikeMinds.

Once LikeMinds became usable, those agents increasingly used LikeMinds to coordinate the continued development of LikeMinds itself. Problems encountered during actual operation became new protocol requirements, which the agents then implemented under my supervision.

So, in the practical engineering sense, LikeMinds became part of the toolchain used to build itself.

It is still an alpha and I don't consider it a hosted service, authorization system, secrets vault or instant-messaging platform. It is a reusable coordination protocol and framework built around a particular idea:

Agents should be able to work together without requiring either an autonomous boss-agent or a human continually carrying messages between isolated conversations.

The public template is here:

github.com/crseabrk/LikeMinds-Template

I'd be particularly interested in feedback from people already using multiple Codex sessions, multiple machines, or several AI coding agents on the same project. I'm curious whether the coordination problems that led me here resemble what others are running into.

[crseabrk]

Read my case study here. 2 flavors.
Scientific Paper style
A-Tale-of-Three-Agents.pdf
Informal Report
A-Tale-of-Three-AgentsEZ.pdf