AXME
Coordinate multiple AI agents through shared intent state — without framework lock-in
Multi-agent systems break at the boundaries: Agent A calls Agent B, B calls an API, C waits for a human — and there's no shared state, no retry, no audit trail across them.
AXME enables agent-to-agent coordination via the AXP protocol.
Coordination is where multi-agent demos die
LangGraph, CrewAI, and AutoGen excel at reasoning inside a graph or crew. Production breaks when Agent A must wait for Agent B on another machine, a human, and an external API — with no shared durable state and no retry when a process restarts mid-handoff.
AXP intents give every participant the same lifecycle vocabulary. Agents submit and wait; humans approve; tools complete — with audit across the chain regardless of which framework produced which step.
Example: research → draft → legal → send
A research agent gathers data (CrewAI), a writer agent drafts (LangGraph), legal reviews (HITL), a sender agent dispatches (OpenAI Agents SDK). Without shared state, handoffs are files in S3 and bespoke message buses per framework pair — AutoGen crews on different hosts cannot talk without you building the broker.
With AXME, each phase is a leg on one intent graph: wait_for_agent between agents, wait_for_human for legal, wait_for_tool for send API — fleet visibility shows the whole chain in Mesh.
Cross-machine AutoGen → AXP
In-process only
# AutoGen crew on one machine OK # different hosts → custom broker
AXP intents
await intent.wait_for_agent("reviewer")
# durable across hostsSOLUTION
How teams solve this with AXME.
Sequential chain
Agent hands off to agent.
Parallel fan-out
Multiple agents, one intent graph.
HITL between agents
Human gate mid-chain.
Cross-framework wiring
Use framework adapters (/integrations/langgraph/, /integrations/crewai/, etc.) to wrap kicks and returns as intent operations. Keep reasoning inside the framework; use AXME for durability, HITL, and handoffs.
Define escalation when a downstream agent fails — intent failure surfaces to ops with full upstream context.
Common questions
- Must all agents use the same framework?
- No — AXP is framework-agnostic. Each agent integrates via SDK or adapter; intents link the chain.
- How do parallel agents join results?
- Model fan-out as child intents or coordinated waits; merge on completion events in your orchestration layer.
- Where is the audit trail?
- Every agent leg and human gate logs to the parent intent — see /audit-trail/.
Related capabilities
Related links
Ship your first durable agent — in under 10 minutes.
Free tier. No credit card. Self-host or hosted — your choice.