Practical Use Cases — ZHD & PROBE
Two end-to-end agentic systems that managers will actually meet. The Zero Human Department for market intelligence shows what governance looks like when no daily human staff sit inside the function. PROBE shows what a structured multi-agent pipeline looks like when every phase transition needs operator authorisation.
- Govern a Zero Human Department without operating it day-to-day.
- Read a PROBE pipeline and identify every phase-boundary intervention point.
- Distinguish Monitor / Verify / Override / Escalate as discrete operator modes.
- Map each use case to the HIMM dimensions it stresses.
Market Intelligence without a market intelligence team
A multi-agent system that scouts signals, synthesises intelligence, evaluates its own outputs, maintains memory across cycles, and delivers a fresh briefing — autonomously. One human operator governs from outside.
01Human Operator · Governance layerverify
You set direction, review outputs, handle exceptions, answer for outcomes.
Verify weekly intelligence reports before distribution; monitor evaluation scores; escalate when quality drops below threshold.
02Orchestrator · Directorverify
Decomposes operator objectives into scout tasks, dispatches in parallel, consolidates results.
Review and approve the operator brief at the start of each cycle. Investigate repeated recycling — it signals drift in synthesis or evaluation.
03Scout agents (Web · News · Regulatory · Social · Competitor)monitor
Parallel specialist scouts that gather signal by domain.
Monitor coverage weekly. Regulatory high-impact findings reach you directly and require verification before synthesis.
04Synthesiser + Self-evaluator + Publisheroverride
Synthesises briefs, scores its own output, publishes only after the evaluator clears it.
Verify the final brief after evaluation — your primary governance touchpoint. Halt publication on threshold failure.
05Memory (Episodic · Semantic · Working)monitor
Shared infrastructure: what happened, what things mean, this cycle's context.
Audit semantic memory quarterly — outdated definitions silently corrupt every downstream brief.
| Mode | Trigger | Action |
|---|---|---|
| Monitor | Self-evaluation scores, weekly | Review aggregate quality. If average drops below your threshold (e.g. 7/10), pause and investigate. |
| Verify | Weekly intelligence report, before distribution | Read the synthesised report. Apply domain judgment: is anything missing, biased, factually suspect? |
| Verify | Regulatory signal flagged high-priority | Verify directly before the synthesiser processes it. Significance is not delegable. |
| Override | Output fails quality threshold or contains material error | Halt publication. Return the brief with a specific correction instruction. Document the error type. |
| Override | Self-evaluator consistently approves low-quality output | Recalibrate scoring criteria. The evaluator is drifting; this is your call, not the system's. |