Chapter 4: Lumen The Mental Model
Four concepts make Lumen predictable. Once you understand them, every workflow makes sense, and nothing surprises you.
Concept 1: Workflows vs. Agents
A workflow is what you run. An agent is what does the work.
Think of a workflow like a menu item at a restaurant. When you order the W1 PMF Discovery, the kitchen (Lumen's Orchestrator) sequences a specific set of specialist agents — EventIQ, SignalMonitor, DiscoveryOS, MarketIQ, DecideWell, RoadMap, HypothesisLab — in the right order, passing the output of each one to the next.
You never call agents directly. You run workflows. The Orchestrator handles the rest.
Each agent is an expert in one thing. EventIQ validates event schemas. SignalMonitor scores PMF. DiscoveryOS builds opportunity trees. None of them try to do what the others do. This keeps outputs focused and makes it easy to understand which agent produced which recommendation.
The agent directory in Chapter 20 lists all 18 agents, their tier, and their purpose. Refer to it when a report section is unclear, and you want to know which agent wrote it.
Concept 2: Context Slots
Agents do not talk to each other directly. They communicate through named "context slots" — a shared object that lives for the duration of a workflow run.
When EventIQ finishes, it writes its output to slots like validated_event_schema and business_model_type. When SignalMonitor starts, it reads those same slots. When SignalMonitor finishes, it writes pmf_score_by_segment. DiscoveryOS reads that slot next.
This has a practical implication: the order matters and the questions are sequential. When Lumen asks you a question early in a workflow, it is populating a slot that several downstream agents will read. A vague answer produces vague outputs across the entire run.
You will not see slot names in normal usage. But when you are writing prompts (Chapter 15), understanding that information flows forward through the workflow helps you give Lumen what it actually needs.
Concept 3: Evidence Quality
Every agent rates its own output. The ratings are HIGH, MEDIUM, and LOW. The Orchestrator takes the lowest individual rating and uses it as the overall grade for the report.
This is not a confidence score about whether the recommendation is correct. It is a data quality score that tells you what the recommendation is based on.
A HIGH rating means: PostHog data present, sufficient sample size (n > 40), recent data (less than 30 days old), and no major gaps.
A MEDIUM rating means: one of those conditions is missing.
A LOW rating means: most or all of those conditions are missing. The recommendation is based on inference, not measurement.
A LOW recommendation is still useful. Lumen tells you what the framework says, given the available data. But it also tells you what data would change the conclusion. Acting on a LOW recommendation without closing those gaps is your call to make — Lumen just makes the gap visible.
Chapter 5 goes deeper into evidence quality and how to improve your ratings over time.
Concept 4: Oversight Gates
Some decisions are too consequential to make automatically. Pricing changes to existing customers. Segment pivots. AI ethics clearances. These get an oversight gate — a mandatory pause where you must respond before the workflow continues.
There are four gate levels.
Level 0 — Automated. No pause. EventIQ and SignalMonitor run without interruption.
Level 1 — Advisory. The recommendation is written to the report. You can override it within 24 hours by typing OVERRIDE: [reason]. If you do nothing, the workflow proceeds.
Level 2 — Approval required. The workflow pauses. You type APPROVE, DECLINE, or MODIFY: [constraint]. Nothing proceeds until you respond. There is a 48-hour timeout.
Level 3 — Governance approval. A named authority must respond. You cannot self-approve. 72-hour timeout. Used exclusively by DataLayer for AI ethics clearances.
When you see a gate, read it carefully. The gate shows you the decision, the options, the evidence quality, and what each option triggers. It is designed to give you enough information to make a real call without digging into the report first.
Chapter 6 covers every gate level in detail, including what to do when a gate times out.
Putting It Together
Here is how all four concepts connect in a single workflow run.
You run /lumen:strategy. The Orchestrator loads the W3 sequence. Agents run in order, each reading context slots written by the one before it. Each agent rates its own evidence quality. The Orchestrator aggregates the ratings and applies the lowest-wins rule. At the North Star definition step, a Level 2 gate pauses the workflow.
You respond with APPROVE. The workflow continues. The final report is graded and annotated with evidence quality at every section.
That is Lumen. Understand those four concepts, and nothing else will confuse you.