Enterprise AI Bootcamp Demo 4

1. A live diagnostic task

planner: model (claude-sonnet-5)manual search: local stubtickets and parts: synthetic training dataMCP 2026-07-28
The agent is given a machine and a symptom, nothing else. It has to work out which tools it needs, in what order, and what to put in their arguments. Watch the second call: its arguments could only be filled from the first call's result.

The task

Unit AC-250-S serial AC25041872 at Immingham Depot: discharge temperature climbs above the trip point under sustained load and the machine shuts down after about 20 minutes at full load. Diagnose it, identify the correct replacement part for this exact variant, and cite the manual procedure.

The answer key — fault F-101, part AC-THE-1006S — is a property of the generated corpus: one variant-specific remedy part exists per fault. The agent never sees it, and the grading below is a string comparison, not a judgement.

Outcome

Checked against the answer keyResult
Fault code identifiedno F-101
Variant-specific part numberno AC-THE-1006S
End to endno
Distinct tools called0 —
Wall clock266 ms
Trace62ccf6392fd4460a

Answer

The run, step by step

Each plan step shows the tool the planner chose and the scores of the runners-up. Each tool step shows exactly what was sent and what came back over MCP.

step 1: finish0.4 ms
planner failed: TypeError: atlas_shared.llm.LLM.complete() got multiple values for keyword argument 'span_name'
answer0.0 ms

Three servers, one task

The three tools live on three separate MCP endpoints with separate identities and separate tool namespaces. The agent's client reached each over HTTP POST with the headers this protocol revision requires.

EndpointServer identityTools
/mcp/ticketsatlas.field-service.ticketsexport_service_records, get_ticket, post_ticket_update, search_tickets
/mcp/partsatlas.field-service.partscheck_availability, find_part, raise_parts_order
/mcp/manualsatlas.field-service.manualssearch_manual
Run executed when you loaded this page. Nothing on this screen is hardcoded; the source is in Demo4/pipeline/.