AI Agent Workflow Guides

These guides turn the FactVerse AI Agent concepts into operating workflows. Each guide explains what data must be ready, which MCP endpoint to start from, which scopes are required, how outputs should be reviewed, and how results should be recorded.

Use the use-case pages for scenario context. Use the workflow guides when an implementation team needs to build, test, or operate an agent workflow.

Guide map

Guide	Start from	Primary output	Review point
Facility Operations	/mcp/base/, Inspector records, Platform asset context	Asset status summary, alarm context, inspection or work-order draft	Operator or supervisor approves action records
Predictive Maintenance	/mcp/pdm/, /mcp/base/, equipment history, signal quality	Health summary, anomaly context, recommended inspection path	Maintenance engineer confirms the work plan
Physical AI	/mcp/base/, Designer scenes, SimReady assets, simulation services	Scenario package, simulation record, robot or process training context	Engineering owner accepts assumptions and validation notes

Workflow contract

Every agent workflow should document the same contract before implementation:

• Goal: the operational decision or engineering task the workflow supports.
• Boundary: tenant, site, asset group, time range, and source systems.
• Inputs: asset records, live or historical signals, documents, scenes, simulation assumptions, and work records.
• Endpoint and scopes: the MCP endpoint, read scopes, compute scopes, and any write scope required for approved actions.
• Output format: summary, evidence table, draft action, scenario package, or validation report.
• Approval path: who can approve a write action, change an operating record, or accept a simulation result.
• Audit record: source references, prompts or task IDs, tool responses, approvals, and final operator feedback.

Data readiness checks

Run these checks before connecting an agent to an operational workflow:

• Asset identifiers are stable across Platform, DFS, Inspector, documents, and external systems.
• Source-system timestamps and data quality flags are visible to the workflow.
• The MCP client discovers tools at runtime instead of relying on hard-coded names.
• Write scopes are separated from read and compute scopes.
• The workflow can return a useful answer when evidence is incomplete, stale, or contradictory.
• Human approval is captured before any action affects an operating system.

Implementation sequence

1. Build a read-only workflow with the narrowest practical scope.
2. Validate source references, missing-data behavior, and operator review format.
3. Add compute tools only after the input data and assumptions are visible.
4. Add draft write actions behind explicit approval and audit controls.
5. Review completed work records and feed accepted corrections back into the knowledge and data layers.

Next steps

• Start with Facility Operations for asset, inspection, alarm, and work-order workflows.
• Use Predictive Maintenance for health, anomaly, and maintenance-planning workflows.
• Use Physical AI for simulation-ready scenes, robot-training context, and process-planning workflows.