MMESHSYSTEMS

MESH AI PLATFORM / DOCUMENTATION

Agent Teams & supervised execution

Coordinate version-pinned specialist Agents using scoped Knowledge, governed execution, and human-reviewed outcomes.

Agent Teams & supervised execution in the actual Mesh AI Portal
Actual authorized Portal capture, 27 October 2026. Coordinator and specialist version selections, explicit dependencies and execution bounds. Cropped above private Knowledge scope identifiers. This is configuration evidence, not proof of complete Team synthesis.

A Team of focused specialists

A Team defines a coordinator and specialists through immutable Agent version pins. The native runtime schedules approved dependencies, bounds execution, retains execution evidence and checks permissions. Current supervised Team delegation follows an explicit graph; an LLM does not freely choose arbitrary specialists.

What has been validated

  1. Create and pin coordinator, Knowledge and drafting specialists to approved model profiles.
  2. Retrieve Project-scoped policy evidence and pass findings to downstream steps.
  3. Inspect real model activity and persisted execution attempts rather than relying on example output.
  4. Use structured fact extraction and deterministic Workflow v2 branches for conditions and approval requirements.
  5. Pause for human approval or rejection and produce an unsent, cited draft that reflects the decision.
  6. Fail closed on invalid facts, incomplete model responses and unauthorized cross-Tenant access.

Keep Agent Teams and Workflows distinct

Agent Teams coordinate version-pinned specialists, while Workflow Studio defines validated steps, conditions and native approvals. Both share governed execution infrastructure. In development, combined Agent-and-Workflow scenarios have completed using real model inference, scoped retrieval and human review. The earlier multi-Agent Team synthesis UAT returned truncated output; its completion-quality acceptance remains open.

A representative supervised process

A customer submits a request. One Agent extracts bounded, validated facts; a Knowledge step retrieves an approved policy; typed Workflow conditions select the permitted route; a human decides whether to approve; another Agent prepares a complete draft and source citation. No actual financial transaction or external message is automatically sent.

Availability and roadmap

The capabilities described above were validated in a controlled development environment, not a generally available hosted service. Dynamic LLM-directed delegation, live Web Search, email connectors, long-running session recovery, dedicated customer node pools and production autoscaling remain future or deployment-scoped capabilities.