Multi-agent business idea analyst

The problem

Ask a single large model to evaluate a business idea and it gives a fluent, confident answer — with no visible seam where it should have said "this depends" or "I'm not sure." Founders take that fluency as certainty, and the actual uncertainty in market sizing, competitive risk and execution difficulty never surfaces.

The constraint that shaped everything

The obvious fix — one better prompt — doesn't produce disagreement, and disagreement is exactly what a real advisory panel would give you.

Four agents disagreeing is the feature, not a bug.

So the system runs four agents in parallel, each with a narrow mandate and no visibility into the others' output until synthesis. Conflicting conclusions are expected, surfaced, and resolved explicitly rather than smoothed away by a single model's tone.

How it works

# graph definition — four independent nodes fan in to one synthesis node
graph = {
    "research": {"depends_on": []},
    "analysis": {"depends_on": []},
    "risk":     {"depends_on": []},
    "writing":  {"depends_on": []},
    "synthesis": {"depends_on": ["research", "analysis", "risk", "writing"]}
}

Each of the four agents runs against the same idea brief but a different mandate and toolset, with no shared context — deliberately, so one agent's phrasing cannot bleed into another's conclusion. Synthesis reads all four outputs, flags where two disagree on a material point, and states the resolution rather than averaging it away.

The hard parts

Conflict resolution at synthesis. When the risk agent flags a concern the analysis agent didn't weigh, synthesis has to state the disagreement and a reasoned resolution — not silently pick one side.

Cost per run. Four agent calls plus a synthesis call is roughly five times the tokens of a single prompt; keeping each agent's mandate narrow is what keeps the run affordable.

Stopping agents converging on each other's language. Without isolation, agents run in sequence tend to echo the previous agent's phrasing and lose their distinct perspective — running them independently and merging only at synthesis prevents that.

Outcome

The output reads as four distinct voices reconciled into one recommendation, with the disagreements kept visible rather than smoothed into false consensus — which is the actual value of a panel over a single model.