Workload-level provenance metadata for AX agents — LAR-1 integration example

[cloudiaspecula]

AX provides the lifecycle and isolation boundary for autonomous agent workloads. LAR-1 is an optional semantic overlay for describing the provenance and epistemic status of messages produced inside those workloads.

The proposed integration is deliberately workload-level and requires no AX control-plane changes:

1. An AX Workspace checks out or provisions a LAR-1 implementation.
2. The workspace wires an LAR-1-aware MCP server.
3. The agent propagates LAR-1 metadata through MCP _meta and its own messages/results.
4. AX continues to schedule, isolate, observe, suspend, and resume the workload without interpreting the semantic fields.

LAR-1 adds fields such as:

• C — cognition (obs, hyp, inf, rev);
• E — evidence (direct, derived, aggregated);
• L — likelihood;
• V — verification status;
• T / S — temporal and spatial situatedness.

This keeps the layers separate: AX describes infrastructure state, while LAR-1 describes message provenance. In particular, LAR-1 should not be added to TaskStatus, since “the task is running” and “the agent’s claim is verified” are different kinds of state.

We prepared a minimal companion example here:

https://github.com/carlsonchik/larone/tree/main/examples/ax-lar1

The example includes an AX multi-document manifest and a short proposal for a possible future annotation-based integration.

Questions:

1. Does AX intend to support or preserve arbitrary resource annotations in v1alpha1?
2. Would a workload-level integration example be useful before adding AX API surface?
3. Is there a recommended mechanism for declaring agent message metadata or MCP result middleware?
4. Would maintainers prefer a companion example in AX, or a link to an external integration repository?