Parallel POCs: A Better Way to Find Real Requirements
I think companies stuck in that linear ‘analysis → build’ mindset, which is basically a mini-waterfall. A better approach, especially for new problems, might de-risk the long ‘analysis’ phase in agile spirit. We can do that by evolving several small LLMs based POCs in parallel and iterating based on which ones actually solve the problem.
LLMs can totally build green-field POCs, that’s not the issue. The blocker is the requirements… they’re just hidden everywhere in Slack, Zoom calls, emails, and presentations.
So the idea, it’s kinda like a genetic algorithm.
First, you collect all the context from chats and meetings . This is honestly the hard part (Signal vs. Noise), you could build whole startups just to solve this. You’d have an agent capture all the threads, notes, and decisions, keep a source trail, and flag anything that’s still an open item. It’s needs to be fully automated. I built some tools to gather this context, but operating them manually is pointless.
Then, you map the conversation, figuring out the people, their goals, and the limits. The agent would build a graph (neo4j) of all the actors, their intents, and the constraints, and just update it whenever new facts come in.
From there, you create many different requirement versions. Just small changes, trying out different trade-offs. The agent would generate all these variants, tagging the risks and assumptions for each.
Next, you build a ton of POCs in parallel with an agents, one for each of those slightly different requirement choices. The agent scaffolds the code, wires up the integrations, and runs basic tests on all of them.
Once they’re built, you auto-score them based on tests, speed, cost, security, and how well they integrate. The agent runs benchmarks, does static analysis, and publishes a scorecard.
You pick the best few, mix their best ideas (like “crossover”), and run a whole new round.
Finally, the humans step in to decide if the result actually fits the stakeholders, the company politics, and the real-world risks. That’s the final judgment call then implementation.
Looking ahead at next-gen integrations, I really don’t believe AGI is around the corner. The near future is all about getting the right data to power agentic systems. In this flow, the key data is requirements. Context is a new King. This whole context extraction from company data is a huge topic.