You've probably used AI already. You've asked ChatGPT a question and gotten a good answer. Your organization might use Copilot for email and documents. These are useful tools.
But notice the pattern: you ask one question, you get one answer. The answer is well-formulated, coherent, and convincing. It feels right.
And that's exactly the problem — for decisions that truly matter.
One AI voice gives you consensus. Swarm gives you structured dissent — perspectives that actively challenge each other, so you see what you didn't know you weren't seeing.
How it works
1
You describe the problem
You formulate the challenge in your own words. Based on that, the AI designs concrete cases — the specific angles the swarm will attack. You can adjust, add, or remove cases before launch.
When you describe your challenge, the AI designs a set of cases — concrete angles the swarm can tackle. If your overarching question is "Should we enter market X?", cases might become: "What are the real barriers to entry?", "Which competency gaps do we have?", "What are the exit costs if it fails?"
Each case is specific enough for instances to deliver actionable analysis, yet broad enough for different seeds to produce genuinely different answers. A good swarm project typically has 5–10 cases.
You can provide example cases as part of the problem description, but the AI will often suggest angles you hadn't considered. You have full control: adjust, add, or remove cases before the swarm starts.
2
The swarm launches with frequency seeds
9–20 AI instances start in parallel. Each gets a unique combination of five weighted words — a seed — that colors its perspective without closing it. The result: genuinely different perspectives, not the same answer repeated 20 times.
Each swarm project has a set of dimensions — independent axes along which perspectives vary. These can be role (engineer, investor, user), tone (direct, exploratory), focus (risk, opportunity, compliance), and similar. The AI designs the dimensions based on the challenge.
Within each dimension are 5–15 words with different weights (0.03–0.85). High weight means core focus. Low weight is the rare reframer that prevents groupthink. Each instance draws one word per dimension — and the combination becomes the instance's frequency seed.
Example: An instance with the seed "skeptic · direct · honest · leader · compact" will attack the problem completely differently than one with "educator · poetic · technical · curious · expansive". Same question, genuinely different analyses.
Dimensions and seeds are visible in the project and can be adjusted — they are design decisions, not hidden magic.
Both are ways to get AI to produce diverse answers — but they work completely differently.
When role perspective is what matters ("what does a leader vs. a union representative think?")
When the analysis should read like a narrative
When you're discussing leadership teams, stakeholder analysis, concrete role design
When frequency seeds win
When you need genuine divergence, not stereotyped voices
When you want to avoid implicit assumptions about who's "the right person to answer"
When you want transparency in what influences each instance
When you need to stress-test a decision from angles you didn't think of in advance
Why swarm.ai uses seeds
We build swarms to discover what we didn't know to ask about. Personas assume we know which roles are relevant; seeds let the combination itself reveal something new. "What would a poetic-legal-urgent analysis say?" has no named persona — but it often sees things nine real experts overlook.
Opposite of conventional thinking — devil's advocate
poetic
0.04
Metaphoric, visual, symbolic understanding
Each instance draws one word from this dimension — weighted so "strategic" appears far more often than "poetic". Combined with one word from four other dimensions, this forms the instance's unique lens.
A flight is one run of the swarm against one case. All instances start simultaneously, receive the same case and context, but see it through their unique seed. They work independently — no instance knows what the others are answering.
Each instance delivers its analysis. Afterwards, a debrief runs that compiles all answers: where is there consensus? Where do perspectives diverge? Which blind spots were uncovered? The result is a structured report where every finding is linked to the seeds that generated it.
A project with 8 cases can run 8 flights — one per case. You can also run multiple flights on the same case with adjusted seeds, for example to dig deeper into a blind spot that emerged.
3
You read the debrief and see the patterns
Each flight produces a debrief — a compilation of all perspectives. You read it and see: where do the instances agree? Where do they split? Divergence is not noise — it's a signal about blind spots worth attention. You can run new flights to dig deeper into what emerges.
A debrief compiles all instances' analyses into one document. You see each instance's output with its seed visible, followed by a pattern analysis: which findings are consistent across perspectives (solid ground), and which are only seen by one or two instances (potential blind spots).
Each finding is tagged with type (architecture, risk, recommendation, blind spot) and linked to the seeds that generated it — so you can assess why a perspective emerged, not just that it emerged.
Start with the divergence. Agreement is comfortable — it confirms what you already sense. What's interesting is where the instances split. Look at which seeds deviate: is it a skeptic seeing risk the others miss? A technologist spotting an implementation barrier?
Assess relevance. Not all divergence is equally important. Some deviations stem from a seed being less relevant to your context. Others reveal genuine blind spots. Your job is to separate signal from noise — and the debrief gives you the tools to do so.
Iterate. See a blind spot you want to understand better? Run a new flight with adjusted seeds or a more focused case. The swarm is an iterative tool, not a one-off report.
4
You get a structured decision foundation
Not a report — a mapping. Here is where it's robust. Here is where the instances see differently. Every finding is linked to its sources. You sit with enough to decide, not with persuasion toward one conclusion.
Swarm vs. conventional AI use
This is not criticism of the tools you already use. They are good at what they do. Swarm solves a different problem.
Conventional AI use
Swarm analysis
One conversation, one perspective
9–20 perspectives simultaneously
Converges toward the safe answer
Designed to find disagreement
Confirms what you already believe
Challenges what you haven't considered
Chat log — hard to reference
Structured report with traceability
Generalist — knows a bit about everything
Calibrated to your industry and role
Good for daily tasks
Built for consequential decisions
Who is this for?
Swarm analysis is for people who make decisions under uncertainty — where the answer key doesn't exist, where the stakes are high, and where missing something costs more than the analysis.
It's the program director in defense steering a billion-dollar program. The CTO evaluating a platform change. The health authority CIO introducing AI in an environment where errors have patient consequences. The strategy director assessing an acquisition.
In the persona library you can find your role and see what a swarm analysis would uncover — without sharing anything about yourself.
What it isn't
Swarm is not a replacement for the AI tools you use daily. It's not a chatbot, not a copilot, not a writing assistant. It's an analysis system for the moments that truly matter — when you need to know what you're not seeing, not just get confirmation of what you already believe.
Think of it as the difference between asking one advisor versus putting ten experts with different perspectives around the table. Both are useful. But for different things.
Curious?
See if you recognize yourself in the persona library, or reach out directly.