IOT Manager

Role

The IOT Manager is the central coordinator of the BWTS IOT team. It never performs specialist work itself — no database queries, no manual-cause lookups, no email composition. Its sole job is to orchestrate the right agents in the right order, synthesise their outputs, and produce a unified root-cause assessment.

Note

Think of the IOT Manager as a senior engineer in a control room: it reads every specialist’s report, cross-references the findings, and writes the final incident summary — but it never touches the instruments directly.

---

What It Does

1. Receive Alert

   The Alert Monitor creates a task addressed to the IOT Manager containing the structured alert payload (alert_monitor_v1 schema). The Manager parses the alert, identifies the vessel, equipment tag, and parameter in violation.

2. Dispatch Phase 1 — Four Parallel Investigations

   The Manager creates four tasks simultaneously, one for each specialist:

   Task	Assigned To	Purpose
   Trend analysis	Data Agent	Retrieve 7-day sensor trends for the flagged parameter
   Manual-cause search	Manual Agent	Search equipment manuals for known causes of this alarm type
   Maintenance history	PMS Agent	Pull recent maintenance and work-order records for the equipment
   Case history	Casefile Agent	Find past incidents on the same equipment or alarm code

   All four tasks are created with no blockedByIssueIds, so they run in parallel.

3. Wait for Phase 1 Completion

   The Manager monitors task statuses. It proceeds to synthesis only when all four Phase 1 tasks are marked complete. If any task fails, the Manager follows the error-handling protocol described below.

4. Synthesis — Cross-Reference and Score

   The Manager reads every Phase 1 result and applies its synthesis logic to produce:

   • A ranked list of probable root causes
   • A confidence rating for each cause
   • Supporting evidence mapped back to each specialist’s findings

5. Dispatch Phase 2 — Report Generation

   The Manager creates a single task for the Report Agent containing the full synthesis payload. This task uses blockedByIssueIds referencing all Phase 1 tasks to guarantee ordering. The Report Agent formats the final HTML report and delivers it.

6. Close the Loop

   Once the Report Agent confirms delivery, the Manager marks the original alert task as resolved and writes a summary to Para Memory for future reference.

---

Synthesis Logic

Synthesis is the Manager’s core intellectual contribution. It follows a structured cross-referencing process.

Cross-Referencing Matrix

The Manager builds a matrix where rows are candidate causes and columns are evidence sources:

	Data Agent (Trends)	Manual Agent (Known Causes)	PMS Agent (Maintenance)	Casefile Agent (History)
Cause A	Trend spike matches	Listed in manual	Recent part replacement	Similar incident 3 months ago
Cause B	No correlation	Listed in manual	No related work orders	No prior cases
Cause C	Gradual drift visible	Not in manual	Overdue calibration	One prior case

Confidence Rating

The Manager assigns a confidence level based on how many independent sources corroborate a candidate cause:

Rating	Sources Agreeing	Interpretation
HIGH	3 or more	Strong convergence — treat as probable root cause
MEDIUM	2	Plausible — include in report with caveats
LOW	1	Speculative — mention only if no higher-rated causes exist

Tip

When two causes are rated equally, the Manager ranks the one with Data Agent support higher, because sensor trends provide the most objective evidence.

Conflict Resolution

When specialists contradict each other (e.g., the Manual Agent says “sensor drift” but trends show a clean step-change), the Manager:

1. Notes the conflict explicitly in the synthesis.
2. Weights the data-driven source (Data Agent, PMS Agent) over the text-driven source (Manual Agent, Casefile Agent).
3. Flags the conflict for human review in the final report.

---

Team Interaction

graph LR
    AM["Alert Monitor\nEntry point"] -->|"Creates alert task"| MGR["IOT Manager\nCoordinator"]

    MGR -->|"Phase 1 (parallel)"| DA["Data Agent\nSensor trends"]
    MGR -->|"Phase 1 (parallel)"| MA["Manual Agent\nKnown causes"]
    MGR -->|"Phase 1 (parallel)"| PMS["PMS Agent\nMaintenance records"]
    MGR -->|"Phase 1 (parallel)"| CA["Casefile Agent\nPast incidents"]

    DA -->|"Results"| MGR
    MA -->|"Results"| MGR
    PMS -->|"Results"| MGR
    CA -->|"Results"| MGR

    MGR -->|"Phase 2 (after synthesis)"| RA["Report Agent\nHTML report & delivery"]
    RA -->|"Delivery confirmed"| MGR

    style MGR fill:#1a1a2e,stroke:#e94560,stroke-width:3px,color:#fff
    style AM fill:#16213e,stroke:#0f3460,stroke-width:2px,color:#fff
    style DA fill:#16213e,stroke:#0f3460,stroke-width:2px,color:#fff
    style MA fill:#16213e,stroke:#0f3460,stroke-width:2px,color:#fff
    style PMS fill:#16213e,stroke:#0f3460,stroke-width:2px,color:#fff
    style CA fill:#16213e,stroke:#0f3460,stroke-width:2px,color:#fff
    style RA fill:#16213e,stroke:#0f3460,stroke-width:2px,color:#fff

---

Delegation Table

Data Agent

Receives: Equipment tag, parameter name, time window.

Returns: 7-day trend data, anomaly timestamps, statistical summary.

When: Every alert — always dispatched in Phase 1.

Manual Agent

Receives: Alarm code, equipment type.

Returns: List of documented causes from equipment manuals, recommended actions.

When: Every alert — always dispatched in Phase 1.

PMS Agent

Receives: Equipment tag, date range.

Returns: Recent work orders, overdue maintenance items, part-replacement history.

When: Every alert — always dispatched in Phase 1.

Casefile Agent

Receives: Equipment tag, alarm code.

Returns: Past incidents, their root causes, and resolutions.

When: Every alert — always dispatched in Phase 1.

Report Agent

Receives: Full synthesis payload (causes, confidence, evidence map, recommended actions).

Returns: Confirmation that the formatted HTML report was emailed to stakeholders.

When: Phase 2 only — after synthesis is complete.

---

What It Does NOT Do

Caution

The IOT Manager must never bypass the specialist layer. If you see the Manager performing any of the actions below, the workflow is misconfigured.

Forbidden Action	Responsible Agent Instead
Run database queries or fetch sensor data	Data Agent
Search equipment manuals or PDFs	Manual Agent
Look up maintenance records or work orders	PMS Agent
Search past incident case files	Casefile Agent
Compose or send emails	Report Agent
Generate HTML reports	Report Agent
Poll for new alerts	Alert Monitor

---

Error Handling

Specialist Task Fails

If a Phase 1 specialist task fails (times out, errors, or returns empty):

1. The Manager retries once after a short delay.
2. If the retry also fails, the Manager proceeds with the remaining specialists’ data.
3. The synthesis explicitly notes the missing data source and downgrades confidence ratings accordingly.
4. The final report includes a “Data Gaps” section listing which specialist was unavailable.

Synthesis Is Inconclusive

When no candidate cause reaches MEDIUM or HIGH confidence:

1. The Manager marks the synthesis as INCONCLUSIVE.
2. The report is still generated but flagged with a prominent “Requires Manual Review” banner.
3. The Manager writes a note to Para Memory so the pattern can be reviewed during weekly synthesis.

Multiple Alerts — Same Equipment

When a new alert arrives for equipment that already has an active investigation:

1. The Manager checks Para Memory for an in-progress task on the same equipment tag.
2. If found, the new alert is appended to the existing investigation rather than spawning a duplicate pipeline.
3. Phase 1 specialists are notified of the additional alert data via task comments.

---

Tools

Tool	Purpose
Task orchestration	Create, monitor, and close tasks for each specialist via the LifeOSAI task system
Gate dependencies (blockedByIssueIds)	Enforce Phase 1 → Phase 2 ordering so the Report Agent never runs before synthesis
Para Memory	Store investigation summaries, track in-progress equipment, and feed weekly synthesis

Note

The Manager’s effectiveness depends entirely on the quality of its delegation. A well-structured task with precise inputs produces a fast, accurate specialist response. Vague tasks produce vague results.