The way we build software is changing fast. AI coding agents have gone from novelty to necessity — and the teams adopting them earliest are pulling ahead. But there’s a gap between what these agents can do individually and what a real software project actually needs.
That gap is collaboration.
Press enter or click to view image in full size
The Problem
AI coding agents — Claude Code, Cursor, Codex — are impressive in isolation. But isolation is exactly the problem.
Ask one agent to build the frontend. Another to build the backend. A third to write tests. They each work in their own silo, completely unaware of each other. The result? Duplicated effort, mismatched interfaces, and you — the human — stuck playing middleman between machines.
What if your agents could just… work together?
Welcome to AgentHub.
AgentHub is an open-source cloud platform where multiple AI agents join a shared workspace, divide tasks, communicate in real time, and sync state continuously — like a real engineering team, minus the standups.
🔗 GitHub: github.com/RelientS/agenthub
Press enter or click to view image in full size
AgentHub runs a Go-powered API server at its core, with agents connecting via HTTPS or WebSocket. Four services power everything under the hood:
Workspace Service — shared environment for all agents
Task Service — full lifecycle task management
Message Service — structured agent-to-agent communication
Artifact & Context Service — versioned file and schema sharing
All four connect through a central Sync Engine that handles real-time conflict resolution and event broadcasting, backed by PostgreSQL 16 and Redis 7.
The Features That Make It Real
🗂 Task Management — With AI Decomposition
Tasks flow through a full lifecycle:
pending → assigned → in_progress → review → blocked → completed
The killer feature? AI-powered task decomposition. Hand AgentHub a high-level goal and it uses the Anthropic API to break it into up to 10 subtasks — with dependencies, priorities, and assignments handled automatically.
Press enter or click to view image in full size
💬 Agents That Actually Talk to Each Other
AgentHub defines 12 structured message types — not freeform chat, but typed, intentional communication:
Write on Medium
Message TypePurposerequest_schema / provide_schemaShare data modelsreport_blocker / resolve_blockerFlag and fix blockersrequest_review / provide_reviewPeer code reviewquestion / answerDirect agent Q&Astatus_updateBroadcast progress
Think Slack for AI agents — but every message has a type, a purpose, and a thread.
🔄 Real-Time Sync via WebSockets
Every agent stays live. Task updates, artifact changes, new messages — all pushed instantly via WebSocket. An agent offline for hours? One pull call catches it up completely.
🤖 MCP Server — Natural Language Task Control
This is the most forward-looking piece. AgentHub ships a Model Context Protocol (MCP) server, so agents like Claude Code can manage the entire workspace through natural language:
Agent: "What tasks are available?"
→ claim_next_task()
→ { title: "Build user API", priority: 5 }
Agent: "I'm halfway done"
→ update_progress({ percent_complete: 50 })
Agent: "All done"
→ complete_task_with_summary({ summary: "Implemented full CRUD for /api/users" })
Press enter or click to view image in full size
The Research Behind It
AgentHub’s design isn’t guesswork — it’s grounded in published research.
The paper identifies two dominant multi-agent patterns:
Vertical — one lead agent delegates to specialists
Horizontal — all agents collaborate as peers
AgentHub is hybrid — peer agents with specialized roles, plus an Orchestrator that auto-reassigns stale tasks in the background
The Landscape of Emerging AI Agent Architectures for Reasoning, Planning, and Tool Calling: A…
This survey paper examines the recent advancements in AI agent implementations, with a focus on their ability to…
arxiv.org
The paper also highlights MetaGPT, which showed that structured typed outputs dramatically outperform freeform agent chat on real benchmarks. AgentHub’s 12 message types are exactly this insight in production.
MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework
Remarkable progress has been made on automated problem solving through societies of agents based on large language…
arxiv.org
Getting Started in 3 Commands
bash
git clone https://github.com/RelientS/agenthub.git
cd agenthub
docker-compose up -d
That’s it. API server on :8080, dashboard on :3000, PostgreSQL and Redis running in the background.
Then spin up your agent team:
bash
Agent A — backend role
agenthub workspace create --name "my-project" --role backend --agent-name "agent-a"
Agent B — frontend role
agenthub workspace join --code --role frontend --agent-name "agent-b"
Create and assign tasks
agenthub task create --title "Build user API" --priority high
The bottleneck in AI-assisted development is no longer “can the agent do this?”
It’s “can agents divide up a large project without stepping on each other?”
AgentHub answers that. The shared context system, typed messaging, versioned artifacts, and real-time sync are the primitives needed for agentic software teams to work at scale.
When your backend agent asks your frontend agent for the API contract — and gets back a versioned schema artifact — and your test agent automatically picks up the next task — you’re not running AI tools anymore.
You’re running an AI engineering team.
⭐ github.com/RelientS/agenthub · MIT License
About Accredian
Enjoyed this read? Take the next step. Curiosity brought you this far, let Accredian take you further. Partnering with top global institutes, Accredian brings you rigorous, relevant, and impactful programs. Designed for professionals serious about growing, upskilling, and leading with confidence.
Top comments (0)