Calling All Developers
Join us to build Meta Harness — the open, universal orchestration layer for AI agents
We have spent the last four years watching the agent harness evolve from a thin prompt wrapper into the most consequential layer in the AI stack. We have seen frameworks explode, protocols emerge, and entire categories of software get reinvented around agentic AI.
But here is the honest truth: no one has built the harness the world actually needs yet.
The harnesses we have today are good. Some are excellent. But they are still fragmented. Still siloed. Still tied to individual vendors, individual ecosystems, individual visions of what an agent should be. The future demands something bigger — an open, community-driven Meta Harness that belongs to everyone and works with everything.
This is a call to action. We are building it. And we need you.
Why Now?
The rise of agent harnesses documented four distinct phases: prompt wrappers, the framework explosion, integrated harnesses, and the protocol era. Each phase solved real problems but also left gaps.
The Gaps That Remain
Fragmentation
50+ harness frameworks, each with its own tool format, memory system, and orchestration model — none fully interoperable
Vendor Lock-In
Most production harnesses are tied to a single model provider, making it painful to switch or combine models
Inconsistent Safety
Safety implementations vary wildly — from rigorous permission systems to no guardrails at all
Memory Silos
Agent memory is locked inside individual harnesses with no way to share or migrate knowledge
Solo Agents
Most harnesses still think in terms of a single agent, with multi-agent coordination bolted on as an afterthought
No Federation
Agents from different organisations cannot discover or collaborate with each other
These are not minor inconveniences. They are structural barriers preventing AI agents from reaching their full potential. And they will not be solved by any single company — because the solution must be open, neutral, and collectively governed.
The time is right because the building blocks exist. MCP gives us a universal tool protocol. A million-token context windows give us room to work. The patterns for safety, memory, and multi-agent orchestration have been proven in production. What is missing is the community effort to weave them into a coherent whole.
The Vision for Meta Harness
Meta Harness is not a framework. It is a specification and reference implementation for the universal agent orchestration layer. Here is what it aims to be:
Core Principles
Fully Open Source
MIT licensed. No contributor licence agreements that transfer rights. No open-core bait-and-switch. Open means open.
Model Agnostic
Works with Claude, GPT, Gemini, Llama, Mistral, and any future model. Swap providers with a config change, not a rewrite.
Protocol Native
Built on MCP from the ground up. Every tool, every data source, every agent speaks the same protocol.
Safety First
Permission tiers, sandboxing, audit trails, and policy-as-code are not optional modules — they are core architecture.
Universal Memory
A shared memory specification that lets agent knowledge persist, migrate, and compose across harnesses and sessions.
Multi-Agent Native
Designed from day one for multiple agents to discover, delegate, collaborate, and coordinate.
The Architecture
Meta Harness is organised into five layers — each independently useful, collectively transformative.
The Five Layers of Meta Harness
Layer 1: Universal Model Interface
A clean abstraction over every major model provider. Handles prompt formatting, tool-call parsing, streaming, and model routing. Write your orchestration logic once — it works with any model.
Layer 2: MCP Tool Mesh
A dynamic tool layer built on the Model Context Protocol. Agents discover tools at runtime, invoke them through standardised interfaces, and handle results with typed schemas. No custom integrations needed.
Layer 3: Federated Memory
A persistent, categorised, queryable memory system with a shared specification. Memories are scoped (user, project, team), versioned, and portable. Switch harnesses without losing what your agent has learned.
Layer 4: Safety Fabric
Permissions, sandboxing, audit logging, and policy enforcement baked into the core. Every tool call passes through the safety layer. Every action is logged. Every destructive operation requires explicit approval.
Layer 5: Agent Mesh
The orchestration layer for multi-agent systems. Agents register capabilities, discover each other, delegate tasks, share context, and resolve conflicts — within an organisation or across organisational boundaries.
Design Targets
What We Need From You
Meta Harness is only as strong as its community. We need developers across every specialisation to make this real.
How You Can Contribute
Core Engineers
Build the reference implementation — the model interface, tool mesh, memory layer, safety fabric, and agent mesh
Tool Builders
Create MCP-compatible tools and publish them to the Meta Harness registry — file systems, databases, APIs, cloud services, dev tools
Safety Researchers
Design and stress-test the permission model, sandboxing, and policy enforcement. Red-team the system. Find the holes before users do.
Memory Architects
Solve the hard problems — salience detection, graceful decay, conflict resolution, privacy-preserving memory sharing
Documentation Writers
Write guides, tutorials, and examples that make Meta Harness accessible to developers of every experience level
Community Builders
Moderate discussions, triage issues, mentor newcomers, and help build a culture of collaboration and quality
The Roadmap
We are building in the open, with clear milestones and community input at every stage.
Meta Harness Roadmap
Specification Published
Open RFC for all five layers. Community review period. Architecture finalized based on feedback.
Alpha Release
Reference implementation of Layers 1–3 (Model Interface, Tool Mesh, Federated Memory). Early adopter testing.
Beta Release
Layers 4–5 (Safety Fabric, Agent Mesh) added. Plugin ecosystem launches. First production pilots.
Stable 1.0
Full specification and implementation. Governance model formalised. Long-term support commitment.
Federation
Cross-organisation agent discovery and collaboration goes live. The Meta Harness becomes a network.
Learning from the Past
The history of agent harnesses teaches us hard-won lessons that Meta Harness is built to honour:
Lessons Applied
| Feature | Lesson from History | How Meta Harness Applies It |
|---|---|---|
| AutoGPT showed agents need safety controls | Safety Fabric is a core layer, not an optional plugin | |
| Framework explosion created fragmentation | Single specification with multiple implementations prevents ecosystem splits | |
| Vendor lock-in trapped users | Model-agnostic interface means you are never locked to one provider | |
| Bolted-on memory was unreliable | Federated Memory is architecturally integrated from day one | |
| MCP proved protocols beat custom integrations | MCP is the foundation of the entire tool layer | |
| Multi-agent was always an afterthought | Agent Mesh is a dedicated layer designed for multi-agent from the start |
Who Is Behind This?
Meta Harness is not owned by a single company. It is a community initiative with contributions from independent developers, researchers, and organisations who share a belief that the agent orchestration layer should be open infrastructure.
Governance will follow the model of successful open-source foundations:
- Technical Steering Committee elected by active contributors
- RFC process for all major architectural decisions
- Transparent roadmap driven by community input
- No single company veto — decisions by rough consensus
- Code of conduct that prioritises inclusion and constructive collaboration
The Stakes
This is not just about better developer tools. The choices we make about AI agent infrastructure will shape how AI integrates into society for decades.
Two Possible Futures
The Closed Future
A handful of companies control the dominant harnesses. Agents are locked to proprietary ecosystems. Tool builders must support dozens of incompatible platforms. Innovation concentrates at the top. Most developers are consumers, not creators.
The Open Future
A shared, open harness specification enables any developer to build agents that work with any model, any tool, and any other agent. Innovation is distributed. Safety standards are collectively maintained. The ecosystem compounds.
The difference between these futures is not inevitable. It is a choice. And it is a choice that developers — not executives, not investors, not policymakers — are uniquely positioned to make.
Every open-source contribution to Meta Harness is a vote for the open future.
Get Started Today
Your First Five Minutes
Star the Repository
Find us on GitHub and star the repo. It takes two seconds and signals community interest.
Read the Specification
The draft RFC for all five layers is published. Read it. Annotate it. Open issues for anything unclear or wrong.
Join the Discussion
Our community forum is where architecture debates happen. Introduce yourself. Tell us what you are building and what you need.
Pick a First Issue
Good-first-issue labels mark tasks that are well-scoped for newcomers. Pick one and submit a PR.
Spread the Word
Share this post. Tell a colleague. The more diverse our contributor base, the better Meta Harness will be.
The Community So Far
Ready to Build the Future?
Meta Harness is open, it is real, and it needs you. Join the community today and help us build the orchestration layer that AI agents deserve.
Join Meta HarnessThe model provides the intelligence. The harness provides the agency. The community provides the future. Let us build it together.