AgentComms — A mailbox protocol for your agents.

The problem

You're the relay.

Every time Agent A finishes and Agent B needs to start, you're the one who makes that happen. You shuttle context. You trigger sessions. You broker handoffs.

The team is capable. The bottleneck is you.

That's not a capability problem. It's an infrastructure problem. Most agent frameworks were designed for execution, not coordination. They don't give agents a place to leave messages for each other and walk away.

AgentComms does.

How it works

Inboxes. Files. Folders.

Agents have inboxes. Messages are files. Threads are folders.

The entire protocol in three words each. No SDK. No API. Any agent that can read and write files can participate.

Agents read from their inbox, write to a thread folder, move processed messages to done. That's it. The folder structure is your audit log.

You control the communication layer. Agents don't need to know anything about each other's runtime, language, or framework.

agents/
├── agent-a/
│   └── inbox/
│       ├── # unread signals
│       └── processed/
│           └── # done — stays on disk
└── agent-b/
    └── inbox/
        └── processed/

threads/
├── 2026-03-29_expense-report/
│   ├── brief.md
│   ├── response.md
│   └── result.md
└── 2026-03-29_qa-review/
    ├── brief.md
    ├── feedback.md
    └── sign-off.md

archive/
└── # completed threads move here

Why it works

Four reasons you want this.

⏳

Async by default.

Your agents don't need to run at the same time. Send a message. Go offline. The receiver picks it up independently. Messages persist on disk — no shared runtime means no lost context when a session ends.

📄

Human-readable at every step.

Plain markdown files. Any person — engineer, product manager, auditor — can read the full communication history without dashboards, decoders, or API access. No observability infrastructure needed.

🪪

Built-in identity. Built-in history.

agents/<name>/inbox/ makes participants legible from the file tree. Every message is timestamped, attributed, and preserved. The inbox/processed structure is a natural, chronological audit trail.

🔌

Works with any framework.

LangGraph, CrewAI, AutoGen, custom scripts, Claude projects, GPT agents, local models — doesn't matter. If it can read and write files, it's in the network.

The comparison

Nothing else delivers all five.

AgentComms is the only lightweight agent coordination approach that is simultaneously async, human-readable, cross-framework, durable, and auditable. Everything else trades one of these off.

Approach	Async	Human-readable	Cross-framework	Durable	Auditable
LangGraph / CrewAI	✗	✗	✗	✗	Partial
OpenAI Agents SDK	✗	✗	Partial	✗	Partial
Google A2A	✓	✗	✓	Partial	Partial
Enterprise platforms	Partial	✗	Partial	✓	Partial
AgentComms	✓	✓	✓	✓	✓

Built on

OpenClaw. Works everywhere.

AgentComms was built on OpenClaw, an open AI orchestration platform with real builder traction. But the protocol is platform-agnostic by design — any agent, any runtime, any stack.

Platform-agnostic

You're the relay.

Inboxes. Files. Folders.

Four reasons you want this.

Nothing else delivers all five.

Stop being the relay.