You’ve probably dreamed of it: a single prompt, 19 AI agents fanning out like a research SWAT team, and coming back with 25 deep, original insights. No more hours of Googling, no more synthesizing scattered sources. Just one click, and the machines do the heavy lifting.
I’ve seen that dream. I’ve also seen the reality: agents arguing, repeating each other, and needing a human to step in and break the deadlock. The promise of autonomous deep research is real. But the dirty secret is that scaling the number of agents without strict prompt differentiation doesn’t give you more insights—it just multiplies the same biases and hallucinations.
Take the ‘Banksia’ project on GitHub. It’s a three-layer workflow with 19 parallel agents, designed to research anything you want from a single prompt. It sounds like magic. But read the fine print from the creator: ‘The research and review part may conflict… the manager will try to reconcile them first, and request and wait for human response if they can’t resolve the conflict.’ Wait. You mean the ‘autonomous’ system often needs a human to step in? Yes. And it gets worse: ‘It’s also very easy to just get similar ideas from different agents, if you don’t give each agent different instructions.’
So here’s what’s really happening: you spin up 19 agents, they all crawl the same internet, process the same biases, and produce the same shallow takes. Then they start arguing because their training data gave them slightly different opinions. The manager agent—the one you hoped would be a super-coordinator—becomes a traffic cop, waving its hands and begging for a human judge. The manager agent isn’t the bottleneck—it’s the only thing saving you from a chatbot civil war.
This is the trap: we think more agents equals more intelligence. But multi-agent systems are not a scaling problem. They are an orchestration problem. Every new agent you add increases the surface area for conflict, redundancy, and hallucinations. If you don’t design each agent with a unique, narrow role, you’re just running a high-cost echo chamber. The emotional hook—the dream of effortless research—collides hard with the reality of AI agents arguing in circles while you wait for a response.
So what’s the fix? First, stop treating agents as interchangeable worker bees. Give each agent a distinct persona, a unique source of data, or a specific sub-question to answer. Second, hardwire conflict resolution into the manager’s prompt—not as a fallback, but as a primary function. And third, accept that for now, deep research still needs a human in the loop. Not because the AI is stupid, but because the most valuable insight often comes from the conflict itself, not from a unanimous vote.
The future of AI research isn’t about more agents. It’s about better conflict resolution. The systems that win won’t be the ones that scale the fastest—they’ll be the ones that know when to stop scaling and start reconciling. That’s the real breakthrough nobody’s talking about.
FAQ
Q: Isn't this just a teething problem? Won't better models fix the conflict?
A: No. Better models still produce divergent outputs if they're fed the same data. Conflict is a structural problem of parallel agents, not a quality issue. You need orchestration, not just better LLMs.
Q: What's the practical takeaway for someone building a multi-agent system?
A: Differentiate every agent's prompt ruthlessly. Give each a unique role, source, or sub-goal. And build a manager that can detect and resolve conflicts automatically—using human input only as a last resort, not a first response.
Q: Is the contrarian take that human-in-the-loop is actually a feature, not a bug?
A: Yes. The best insights come from resolution of conflicting perspectives. A system that forces a human to adjudicate is actually forcing deep thinking. The goal shouldn't be zero human intervention—it should be human intervention that adds value.