Give five agents the same bug and you get one investigation, five times
Graph Engineering pattern 6 of 15: the Hypothesis Swarm. Assigned hypotheses, no lateral edges, a falsifying verifier — a graph that optimizes for diversity of reasoning, not throughput.
Give five agents the same hard bug and let them talk, and you will usually get one investigation, five times. The first coherent theory recruits every agent's attention; disconfirming evidence gets explained away. The failure is not effort. It is correlated reasoning — and you fix it with structure, not prompts.
Context for new readers: pattern 6 of 15 in the Graph Engineering series, on designing agent organizations as graphs. Pattern 2 (the Investigator–Fixer Loop) handles ordinary bugs. This pattern is for the defects that resist it — intermittent, environment-dependent, spanning subsystems — where the bottleneck is not effort but imagination.
The topology
Hypothesis A --+
Hypothesis B --+
Problem-> Hypothesis C --+--> Verifier
Hypothesis D --+
Hypothesis N --+
all refuted? -> re-seed the swarm
The problem statement fans out to N agents, with a twist: each agent is explicitly assigned a different hypothesis — a data race, configuration drift, a dependency regression, a resource leak — and instructed to build the strongest case for its own explanation and against it. The agents share no lateral edges. They cannot see each other's reasoning, so they cannot converge on it.
The verifier is not a vote counter
Each agent submits evidence to a verifier, which compares the causal stories against the observed facts and attempts to falsify the leading candidates. Three agents finding their assigned hypothesis "plausible" is not evidence; a causal story that survives falsification is. And if every hypothesis is refuted, that is information — the swarm is re-seeded with hypotheses generated from what has now been ruled out.
Waste, by design
This graph optimizes for diversity of reasoning, not throughput. Half the swarm's work is "wasted" on wrong hypotheses by design — that waste is the price of not spending three days confidently fixing the wrong thing. It is differential diagnosis, drawn as a graph: competing explanations each get a full-strength advocate, and the structure guarantees it.
Where it breaks
- Leaky isolation. A shared scratchpad, a common board, agents reading each other's output — correlation returns through the side door. Isolation is the load-bearing property.
- Verifier as tally. Confidence scores aggregated across agents measure agreement, not truth.
- Hypothesis set too narrow. The swarm can only find causes someone thought to assign. The re-seeding edge exists precisely because exhausting a hypothesis set teaches you where the answer is not.
The dimension answers
| Dimension | Answer |
|---|---|
| Nodes | N hypothesis agents, verifier |
| Edges | Fan-out with assigned hypotheses; evidence fan-in; no lateral edges |
| Cycles | Re-seeding loop when all hypotheses are refuted |
| Humans | Optionally supply or veto the hypothesis set |
| Exit rule | One hypothesis confirmed by evidence and surviving falsification |
The deep point generalizes past debugging: whenever you fan the same question out to N agents and let them share context, you are buying redundancy and calling it coverage. Assignment plus isolation is what turns N agents into N perspectives.
Next: pattern 7, the Architecture Debate Graph — where disagreement is not friction but the product.
Building this with design partners → rysh.ai/design-partner