MMESHSYSTEMS

MESH AI PLATFORM / DOCUMENTATION

Workflow Studio

Design versioned Workflows with typed decisions, Agent steps, scoped Knowledge, native approvals, and inspectable execution.

Workflow Studio in the actual Mesh AI Portal
Actual Workflow Studio screenshot supplied for the website on 27 October 2026. Six connected nodes show Start, coordinator, specialist Agent steps, aggregation and End. Private Team and Agent resource strings are explicitly redacted; the remaining interface and completed badges are unchanged. Recorded step completion does not prove complete model output, dynamic LLM-directed delegation or production readiness.

Build and publish

  1. Create a draft with Start and End nodes.
  2. Add Agent, Knowledge Retrieval, approved REST Tool, Condition, Human Approval, Fan-Out or Join nodes as supported.
  3. Configure authorized resources and connect compatible handles.
  4. Save the draft, validate and publish an immutable version.
  5. Monitor the resulting run and step attempts in Operations.

Canvas controls

Pan and zoom, Fit View, minimap, undo/redo and connection insertion help edit larger graphs. Focus Canvas and independent panels create more room. Presentation metadata such as positions and viewport does not grant runtime permissions.

Validation and parallelism

Server validation rejects invalid graphs, cycles and unauthorized resources. Fan-Out creates independent durable steps; Join follows its configured policy. Workers persist attempts and use leases for eligible recovery. Do not assume exactly-once external side effects.

Test Chat

Each Send starts one independent run of the saved published version. It is not multi-turn Agent Chat memory. Unsaved edits are not executed. For complex input contracts, use Advanced JSON. Accepted, waiting, running, approval-required and completed states have different meanings.

Recovery and approvals

Use run details to inspect actual output, node status, attempts and trace references. Make approval decisions through the authorized inbox. Closing Test Chat does not cancel a run. Cancellation does not undo actions already completed.

Typed decisions and Agent-assisted review

Opt-in Workflow execution v2 supports typed numeric and Boolean conditions, selected true/false branches, persisted skipped steps and authorization checks. Controlled development UAT also exercised real Agent extraction, pinned policy retrieval, human approval/rejection and complete cited draft generation through the native worker. Older Workflow graphs keep legacy behavior.