{"slug": "i-built-a-team-of-ai-agents-to-find-startup-opportunities", "title": "I Built a Team of AI Agents to Find Startup Opportunities", "summary": "A developer built a multi-agent system using Hermes Agent to conduct startup research, deploying four specialized AI agents that research markets, investigate competitors, audit evidence, and rank B2B AI SaaS opportunities. The system, designed to reduce uncertainty rather than generate ideas, produces structured research packages with evidence-backed conclusions.", "body_md": "Most people use AI for startup research like this:\n\n“Give me 10 promising AI startup ideas.”\n\nA few seconds later, you get a polished list.\n\nThe problem?\n\nYou have almost no idea which conclusions are backed by evidence, which are assumptions, and which are simply the model confidently connecting dots.\n\nSo I tried something different.\n\nInstead of asking one AI agent to find startup ideas, I built a small Startup Intelligence team using Hermes Agent.\n\nThe system uses four specialized AI agents that research markets, investigate competitors, audit evidence, challenge each other’s conclusions, and ultimately rank promising B2B AI SaaS opportunities.\n\nAnd rather than producing another Markdown document full of ideas, the workflow produces structured research containing:\n\nHere’s how the system works.\n\nStartup research looks easy until you actually need to decide where to spend your time and money.\n\nA few signals can be surprisingly misleading.\n\nA market receiving hundreds of millions in venture capital doesn’t necessarily mean customers are willing to pay for another product.\n\nEspecially when the numbers come directly from vendors.\n\nA logo doesn’t tell you how much the customer pays, how heavily they use the product, or whether they’ll renew.\n\nBut those companies may target completely different buyers, workflows, or budgets.\n\nGeneric AI research tends to compress all these signals into something like:\n\n“This is a rapidly growing market with strong demand and significant opportunity.”\n\nThat sounds convincing.\n\nBut as a founder, it doesn’t tell me what I actually need to know:\n\nIs this opportunity strong enough to investigate further?\n\nSo instead of optimizing the system for generating ideas, I optimized it for reducing uncertainty.\n\nI created four separate Hermes Agent profiles.\n\nEach agent has:\n\nThe architecture looks roughly like this:\n\n```\n                     ┌─────────────────────┐\n                     │   Startup Director  │\n                     └──────────┬──────────┘\n                                │\n                    Defines scope + rubric\n                                │\n                ┌───────────────┴───────────────┐\n                ▼                               ▼\n      ┌───────────────────┐          ┌─────────────────────┐\n      │ Market Researcher │          │ Competition &       │\n      │                   │          │ Signals Analyst     │\n      └─────────┬─────────┘          └──────────┬──────────┘\n                │                               │\n                └───────────────┬───────────────┘\n                                ▼\n                     ┌─────────────────────┐\n                     │   Skeptic Editor    │\n                     └──────────┬──────────┘\n                                ▼\n                      Evidence-backed\n                     research package\n```\n\nThe important part isn’t simply having four agents.\n\nIt’s giving them different judgment lenses.\n\nLet’s look at each one.\n\nThe Startup Director acts like the project lead.\n\nIts first job isn’t browsing the web.\n\nInstead, it converts an ambiguous founder question into a bounded research problem.\n\nFor example:\n\nFind five promising B2B AI SaaS startup markets in North America with meaningful evidence of customer demand and new company formation between January 2025 and August 2026.\n\nBefore delegating research, the Director creates two important artifacts:\n\n```\nbrief.md\nrubric.md\n```\n\nThe brief defines things such as:\n\nThe rubric defines:\n\nThis matters because otherwise different agents can quietly interpret the same research question differently.\n\nOne might optimize for funding.\n\nAnother might optimize for TAM.\n\nAnother might optimize for how many startups exist.\n\nThe rubric establishes the rules before the evidence is collected.\n\nThe Market Researcher investigates whether there is meaningful evidence that customers actually have the problem.\n\nIt looks for signals including:\n\nBut simply finding a claim isn’t enough.\n\nThe research needs to preserve where that claim came from.\n\nThat distinction becomes extremely important later when another agent audits the research.\n\nThe second researcher approaches the same markets from another direction.\n\nInstead of primarily asking:\n\n“Is there demand?”\n\nIt asks:\n\n“How is this problem being solved today?”\n\nThat means researching:\n\nIt also investigates whether market attention is translating into something stronger:\n\nadoption, payment, retention, or recurring usage.\n\nThis distinction matters.\n\nA market can be extremely popular on X, LinkedIn, Product Hunt, or Hacker News while having surprisingly little evidence that businesses are paying to solve the problem.\n\nThe Market Researcher and Competition Analyst therefore examine overlapping markets but with different objectives.\n\nThis might be the most important agent in the entire workflow.\n\nMost AI pipelines optimize for generating an answer.\n\nThe Skeptic Editor optimizes for finding reasons that answer might be wrong.\n\nIt doesn’t simply read the summaries produced by the other agents.\n\nIt examines the underlying:\n\n```\nresearch files\nclaims ledger\nsource URLs\nsupporting passages\nclaim classifications\ncounter-evidence\n```\n\nThen it asks questions like:\n\nIf a source supports only a weaker statement, the claim gets narrowed.\n\nIf the evidence quality is poor, confidence gets downgraded.\n\nAnd if there isn’t enough evidence?\n\nThe answer can simply become:\n\n```\nUNKNOWN\n```\n\nThat’s a feature, not a failure.\n\nAn AI research system should be able to admit uncertainty.\n\nThere was another architectural problem.\n\nSeparate Hermes profiles don’t share conversation history.\n\nThat means I needed a durable coordination mechanism.\n\nThis is where Hermes Kanban becomes useful.\n\nThe Startup Director creates two independent research tasks:\n\n```\nMarket Research\nCompetition & Signals\n```\n\nBecause neither depends on the other, they can run in parallel.\n\nThe Director then creates another task:\n\n```\nSkeptical Review\n```\n\nBut this task has both research tasks as parents.\n\nConceptually:\n\n```\n             Startup Director\n                    │\n           ┌────────┴────────┐\n           ▼                 ▼\n     Market Research    Competition\n           │                 │\n           └────────┬────────┘\n                    │\n             BOTH COMPLETE\n                    │\n                    ▼\n             Skeptic Editor\n```\n\nThe Skeptic Editor therefore doesn’t start synthesizing conclusions while half the research is still missing.\n\nThis creates an actual dependency graph rather than simply launching several agents and hoping they coordinate.\n\nFor this experiment, I asked the system to compare five specific B2B AI SaaS startup markets in North America.\n\nThe industry scope covered:\n\n🏥 Healthcare & Life Sciences\n\n💻 Software Development\n\n🎧 Customer Support & Contact Centers\n\n📊 Finance & Accounting\n\n🛍️ Retail\n\nI also explicitly excluded:\n\nThe goal wasn’t:\n\n“Which AI industries are growing?”\n\nThat’s too broad to be useful.\n\nThe agents instead compare specific combinations of:\n\n```\nBuyer + Pain + Workflow + AI Advantage\n```\n\nThat’s much closer to the level at which a founder can actually validate an opportunity.\n\nOne requirement fundamentally changed the quality of the output:\n\nEvery important market claim needed to be backed by a source.\n\nThe claims dataset preserves fields such as:\n\n```\nclaim\nstatus\ncompany / market\nclaim type\nsource URL\nsource title\npublication date\nretrieval date\nsource tier\nsupporting passage\n```\n\nThis makes the research auditable.\n\nInstead of seeing:\n\n“Companies are increasingly adopting AI for this workflow.”\n\nYou can inspect the exact evidence that caused the system to make that statement.\n\nAnd that means another agent—or a human—can challenge it.\n\nThe Skeptic Editor produces several structured CSV files.\n\n```\nmarkets.csv\nexperiments.csv\ncompetitors.csv\ncompanies.csv\nclaims.csv\n```\n\nThis makes the results much more useful than one huge research document.\n\nThis contains the ranked opportunities.\n\nFor each market, the dataset can include:\n\nNow I can compare opportunities instead of reading five unrelated research reports.\n\nThis captures the companies discovered during research.\n\nFor example:\n\nThis helps answer:\n\nWho is already attacking this problem, and how?\n\nCompetition isn’t limited to startups.\n\nThe real alternative could be:\n\nUnderstanding the status quo is often more useful than counting startups.\n\nThis is effectively the evidence ledger.\n\nEvery important conclusion can point back to the evidence supporting it.\n\nThat creates a useful chain:\n\n```\nRecommendation\n      ↓\nMarket finding\n      ↓\nClaim\n      ↓\nSource\n      ↓\nSupporting evidence\n```\n\nThis makes the final recommendation much harder to fake with confident language.\n\nFinding an interesting market still doesn’t mean you should build a product.\n\nSo every promising opportunity gets converted into a small validation experiment.\n\nThe experiments can contain:\n\n```\nHypothesis\nTarget buyer\nProcedure\nChannel\nSample size\nCost / effort\nSuccess threshold\nFailure threshold\nDecision rule\nEvidence required\n```\n\nThis changes the final question.\n\nInstead of:\n\n“Should I build this startup?”\n\nThe system asks:\n\n“What’s the cheapest experiment that could prove this opportunity weaker or stronger?”\n\nThat’s a much better founder question.\n\nAfter building this workflow, this was my biggest takeaway.\n\nThe value of multi-agent systems isn’t simply that you can run more AI agents.\n\nFour agents producing four versions of the same answer isn’t necessarily better than one.\n\nThe advantage appears when agents have:\n\ndifferent responsibilities + independent evidence + explicit dependencies + durable state + adversarial review.\n\nThen you’re building something closer to a research process.\n\nThe final system doesn’t claim:\n\n“I found the perfect startup.”\n\nInstead, it tells you:\n\nThat distinction matters.\n\nBecause when you’re deciding where to spend months building a product, a confident answer isn’t enough.\n\nYou need a process that’s harder to fool.", "url": "https://wpnews.pro/news/i-built-a-team-of-ai-agents-to-find-startup-opportunities", "canonical_source": "https://dev.to/vivek_shetye/i-built-a-team-of-ai-agents-to-find-startup-opportunities-3309", "published_at": "2026-08-11 18:37:09+00:00", "updated_at": "2026-08-11 18:49:16.937066+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-tools", "ai-research"], "entities": ["Hermes Agent"], "alternates": {"html": "https://wpnews.pro/news/i-built-a-team-of-ai-agents-to-find-startup-opportunities", "markdown": "https://wpnews.pro/news/i-built-a-team-of-ai-agents-to-find-startup-opportunities.md", "text": "https://wpnews.pro/news/i-built-a-team-of-ai-agents-to-find-startup-opportunities.txt", "jsonld": "https://wpnews.pro/news/i-built-a-team-of-ai-agents-to-find-startup-opportunities.jsonld"}}