🌟 Subagents vs. agent teams
In Module 1.2, you saw the five lenses. What remains is understanding how they run. In STORM, the main session triggers the five lenses as subagents in a single message — and these subagents don’t talk to one another. Each researches independently and returns the result only to the main session. That's different from a agent team.
🟡 New here? — three terms
- Subagent: a working copy of Claude that the main session creates for a specific task (e.g., "be the economist and research X"). It does the task, returns text, and disappears. It has no memory of the other tasks.
- Main session: the conversation where you is there. It dispatches the subagents, receives their results, and assembles the report. The “conductor.”
- Agent team: an arrangement in which several agents talk to each other — they debate and respond to one another. Richer, but more expensive and less predictable. STORM no use this.
The image compares the two arrangements. On the left, a agent team: four agents all connected to each other, talking in a mesh. On the right, STORM: the main session launches isolated subagents that respond only to it, and it’s the one that brings everything together into a result.
On the left, an agent team: the gray lines connect all with all — they debate. On the right, STORM: the arrows only go from the main session to each subagent and back to a single result; the five lenses never connect to each other. STORM uses the right-hand approach because the opinions are cross-referenced afterward, in the contradiction map—not in the middle of the conversation.
✗ Agent team (not what STORM does)
- ✗Agents talk → many messages, higher cost
- ✗One agent "convinces" the other → contaminates their opinions
- ✗More unpredictable and harder-to-audit result
✓ Isolated subagents (STORM)
- ✓Each lens gives its opinion without seeing the others → independent opinions
- ✓Run everything in parallel in a single message → faster
- ✓The main session cross-checks the opinions afterward, in a controlled way
⚠️ Convergence ≠ consensus
This is the method’s most important safeguard, and it’s easy to forget. When the five lenses agree in something, it's tempting to treat that as proof. But the five lenses have the same author: Claude wrote the five prompts and Claude ran the five subagents. The panel is original, not advice from five independent human experts.
🛡️ The safeguard, in one sentence
Agreement across lenses is a strong hypothesis — a good bet for where to look first—, not independent proof. And the STORM report always reveals that the panel was assembled by the author; it never presents convergence as "field consensus."
It’s the evidence behind each lens that gives it weight — not the fact that five roles said the same thing.
✓ What convergence proves
- ✓Whether the point holds up under several different framings
- ✓Whether it belongs at the top of the list to verify first
- ✓Whether it’s a good candidate for a “probably true finding”
✗ What convergence does NOT prove
- ✗Whether the field of real experts agrees
- ✗Whether the cited number/study exists and is correct
- ✗Which five “voices” count as five independent sources
Why this protects you
Without this rule, you’d share a report saying “the five experts agree”—when in practice, you’d be citing the same model five times. Being upfront that the panel has a single author is what keeps you honest. The real proof comes in Phase 4 (topic 6), when each citation is checked against the original source.
🗺️ The contradiction map
After the five perspectives return, the main session assembles the contradiction map. Important detail: this is done internally, without dispatching more agents — the main session reads the five briefings and cross-references everything on its own. The map has five items, and it isn’t a separate deliverable: it’s the raw material where the report’s findings come from.
Direct conflict
Where two or more lenses make claims opposing. This is where each side's specific claim is named—not just “they disagree on the topic,” but “lens A says X, lens B says not-X.”
Strong × weak evidence
Which side is better supported and which is weaker — using the evidence hierarchy (topic 5). Someone who brought a peer-reviewed causal study beats someone who brought an analogy.
The question that resolves it
A single empirical question that, if answered, would resolve the biggest contradiction. This is what separates a "fixed opinion" from a "question you can investigate."
Universal agreement
What all the lenses confirm it — including the skeptic. It’s the strongest candidate for a true finding. You’ll dig deeper in topic 4.
Blind spot
What none covered. It becomes the “missing sixth lens” and informs the boundary question — also in topic 4.
🟡 New here? — “empirical”
A question empirical is one that can be answered with real-world data (“what was the adoption rate in 2024?”), not opinion. Item 3 on the map looks specifically for that question: the one that would measure who is right.
🤝 Universal agreement and blind spot
Two items on the map deserve highlighting because they become central parts of the report. They are the two extremes: what everyone confirms and what no one looked.
🤝 Universal agreement
The point that appears in all the lenses — including the skeptic’s, whose mission was to attack it. If even the one who tried to tear it down couldn’t, it’s your most likely finding.
→ Remember topic 2: even so, this is a strong hypothesis, not proof. It only becomes a conclusion after verification in Phase 4.
🕳️ Blind spot → missing 6th lens
The angle that none touched by one of the five lenses. Instead of ignoring it, STORM names it the “missing sixth lens": what perspective, if it existed, could change the conclusions?"
→ This feeds into the boundary question: the question whose answer would change everything.
What no one saw doesn’t disappear: it becomes the next question to investigate. That’s how the report tells you “where to look next.”
📊 Reliability = evidence quality
Each finding in the report gets a reliability score from 1 to 10. This score doesn't measure how "confident" the model "feels" — it measures the source quality behind the claim. To do this, STORM uses a fixed hierarchy: the evidence ladder, from strongest to weakest.
🪜 The evidence ladder
- 1Peer-reviewed causal — studies showing cause and effect that have passed peer review (e.g., controlled trials, meta-analyses). Stronger.
- 2Official / financial data — government statistics, audited financial statements, public records.
- 3Commissioned single research — a survey or paid study, without replication. Useful, but treat it with caution.
- 4Analogy — “this looks like that thing from the past.” It sheds light, but doesn’t prove anything.
- 5Preprint — published study before peer review. Weaker. Promising, but unverified.
🟡 New here? — reliability ≠ confidence
Reliability is objective: it comes from the source’s position on the ladder above. Subjective confidence is how much something sounds convincing. A text can sound utterly confident while relying on a weak preprint — low score. And it can sound lukewarm while relying on a solid causal study — high score. In STORM, the evidence ladder wins, not the tone.
✅ Verify primary sources
This is Phase 4 — and it’s what separates STORM from “asking for a polished report.” Before delivering, the main session launches verifiers (more subagents, ~4 to 6, one per citation group) that check each citation against the actual source: does the study exist? Is the number correct? Is it published or a preprint? A primary source is the data’s direct source — the study itself, the financial statement, the official document — not a blog that summarized it.
Each citation receives a verdict:
Verificação: 11/11 citações checadas → 2 fabricadas (cortadas), 3 corrigidas, 1 rebaixada
The banner should be true: if 2 numbers didn’t exist, it says “2 fabricated.” Weak or disputed citations are downgraded for the "contested signal" sidebar instead of becoming a conclusion.
✓ Properly verified
- ✓Checks the primary source, not someone else’s summary
- ✓Cut or downgrade what can’t be verified—without disguising it
- ✓The banner reflects what was actually found
✗ Red flags
- ✗Report delivered without Phase 4—it isn’t STORM
- ✗“0 fabricated” banner without actually checking
- ✗An unverified number kept as if it were fact
Self-recovery (optional): a report was delivered without running Phase 4. What does that mean?
📌 Module summary
Next module:
2.1 — Skill Anatomy (starts Track 2: Practice)