Open Standard · EightX Agent Manifest · v0.1 · Apache 2.0 (spec) · MIT (tooling)

Define once.
Run anywhere.
Move freely.

We published EAM because the intelligence economy needs open infrastructure that no single company controls. One declarative YAML file describes your agent's skills, tools, sub-agents, and policies, independently of which provider executes it. Compile-to-Claude ships today; OpenAI, Gemini, Copilot and Agentforce are on the published v0.2 roadmap. The intelligence your agents build stays yours.

View on GitHub → See it compile
Spec - JSON Schema draft 2020-12
License - Apache 2.0 (spec) · MIT (tooling)
Validator - Node.js, CLI + library
Compilers - Claude (live) · OpenAI · Gemini (v0.2)
What it solves

Lock-in is a design choice. EAM unmakes it.

Every model provider ships their own agent runtime - and their own format. Agentforce DX. Copilot Studio. Anthropic's filesystem agents. OpenAI Assistants. They don't talk to each other. They don't audit the same way. They don't bill the same way. And once you've built into one, you've bought into one. Your agents' intelligence - everything they've learned about your business - gets locked inside a vendor's infrastructure. EAM is the exit door.

01

One manifest, many runtimes

Write the agent once in EAM. The manifest declares which runtimes are preferred, supported and degraded, so portability is stated rather than assumed. Compile-to-Claude ships today; the other targets return a structured not-implemented error with a roadmap, never silent best-effort output. Your agent's identity, skills and learned context travel with it.

02

Provider-neutral A2A

A normalized session shape (messages, tool calls, metadata) means a Claude-runtime agent can hand a task to a GPT-runtime agent on the same Conductor. No per-provider branching.

03

Policies travel with the agent

Data residency, PII handling, alignment floor, human-review thresholds, authority limits - declared in the manifest, enforced by every compiler.

90-second demo

Manifest in. Working agent out.

This is the live Cassie agent - agnt8x's own customer-facing assistant. Same manifest, two compile targets. Toggle between them to see what changes (model, system prompt format, tool schema) and what doesn't (the agent's identity, skills, policies).

cassie.eam.yaml

What you're seeing

Try it yourself
$ npm install -g @eightx/eam-validator
$ eam-validate cassie.eam.yaml
✓ valid · 0 errors

$ node compilers/claude/compile.js \
    cassie.eam.yaml --out=./build/
✓ system-prompt.md
✓ mcp-servers.json
✓ subagents.json
✓ compile-report.json
Compiler status

Honest about what ships.

Stub compilers return structured EAM-COMP-NOT-IMPLEMENTED errors with a roadmap - never silent best-effort output. You always know which runtime is live, which is staged, and which is documented but pending.

Target Compiler Runtime adapter BYOK Status
Claude (Anthropic) system prompt + MCP servers + subagent registry Live Live Shipping
OpenAI Roadmap to Assistants API v2 Chat Completions (v0.1) → Assistants (v0.2) Live v0.2
Gemini (Vertex) Roadmap to function calling + tools Stub Format-validated v0.2
Microsoft Copilot Studio Roadmap to Bot Framework Skill Stub OAuth credentials v0.2
Salesforce Agentforce Roadmap to Agent Conversation API Stub JWT bearer v0.2
Governance

Built to be donated.

We published EAM as an open spec because we believe the intelligence economy needs infrastructure that no single company controls. EAM is governed by EightX Labs Ltd today. When adoption justifies it, the spec will be donated to a neutral foundation - the Linux Foundation AAIF is the current target. The marketplace stays at agnt8x. The spec belongs to everyone. That is not a marketing line. It is a legal commitment.

Stop rewriting the same agent five times.

Write it once in EAM. Compile-to-Claude is live; the OpenAI, Gemini, Copilot and Agentforce compilers land in v0.2. Switch models when the economics change. Your agents' intelligence stays yours - it doesn't belong to the runtime.

github.com/eightx-labs/eam-spec Try on agnt8x