Durable execution for AI agents and services.
Submit once. Get result later.
No polling. No webhooks. No retry code.
Open protocol. 5 SDKs.
http://cloud.axme.ai
http://github.com/AxmeAI/axme
AXME
Loading...
Vercel being explicit about agentic features and training controls is good. The harder part starts after disclosure: what actions can the platform take, when does a human need to approve, and how do you resume safely later. That plumbing gets real fast.
The policy crowd is finally staring at the part builders usually skip: not model evals, org evals. Who gets slowed down, who gets audited, who owns failure, who has to sign off. Enterprise agents will live or die on that plumbing, not benchmark charts.
Temporal Replay's agenda getting shared now makes sense. The useful bit is not conference hype, it's that durable execution, tool routing, observability, and real HITL are finally being treated as the base layer for agents. That's the AXME problem too.
Webhook retries are a tax on every agentic workflow. Backoff, jitter, dedupe, DLQ, signatures, idempotent tool call logic - all for "tell me when it's done." Durable execution should own delivery. App code should own the work.
AgentRFC is hitting a real nerve. Multi-agent protocols need more than message formats, they need clear trust boundaries, auth, replay rules, and conformance. That also shows why AXME cares so much about durable intent delivery and HITL across agent handoffs.
Temporal centering Durable Execution for agentic AI feels right. Most agent failures are boring backend failures: lost state, duplicate steps, orphaned approvals, handoffs that never land. That is the same hole AXME is trying to close with intent delivery and real HITL.
Your agent asked for approval at 11:52. Reviewer went to lunch. By 3pm the process is gone, the session expired, and nobody knows what state the agentic workflow is in. HITL is still the part most stacks fake.
You can spend 2026 wiring webhook retries into every agentic workflow, or treat delivery like infrastructure. at-least-once, timeout, escalation policy, resumable execution. App code should do the work, not babysit the callback.
The quiet win here is putting Responses traffic behind real serving controls. Tool calls get weird in production fast - retries, idempotency, partial failures, schema drift. A single endpoint helps, but somebody still has to own state when the model says "continue.
ServiceNow making MCP approval state enforceable is the right move. Governance has to block tool use, not just file a warning. But once a server is paused or approval is missing, the agent flow still needs durable wait states, reassignment, and resume. That's the AXME layer.
