What Inspired Us
Every year, billions of dollars in VC funding go to a tiny fraction of startups. Not because the others are bad. Because nobody had time to look at them.
A VC analyst can evaluate around 200 companies per year. There are over 500,000 startups launched every year. That gap is not a funding problem. It is a time problem.
We asked one question: what if an autonomous agent could do what an analyst does, but at machine speed, with cited sources, and for $1?
That question became DealAgent.
What It Does
DealAgent compresses 20 hours of VC due diligence into 4 minutes. Five specialized agents coordinate, pay each other, and publish the output to the open web with zero human involvement.
The pipeline:
- Research Agent uses Nimble to fire 4 parallel web searches simultaneously, gathering live signals on founders, traction, market size, and risk
- Scoring Agent uses Claude to turn that research into structured 0 to 10 scores, every score backed by a real source URL from the web
- Benchmark Agent queries ClickHouse against every deal ever scored, comparing this company to real historical patterns in milliseconds
- Publisher Agent uses Senso to publish a grounded cited report to cited.md, live on the agent readable web
- Payment Agent settles 1 USDC from the requesting agent on Base via x402, machine to machine, no human touches billing
Every decision across all five agents is traced in Datadog LLM Observability.
How We Built It
Backend: Python and FastAPI with async throughout. All five agents are async functions coordinated by a central pipeline.
Web intelligence: Nimble powers the Research Agent with four parallel searches using asyncio.gather. This is what makes the research fast and grounded in real sources.
Analytics: ClickHouse Cloud stores every report ever generated. When a new company comes in, ClickHouse runs a benchmark query:
SELECT count(), avg(overall_score), countIf(overall_score >= 7.0) FROM dd_reports
This makes the scoring trustworthy. Not just an LLM opinion but a score calibrated against real historical data.
Publishing: Senso publishes every report to cited.md as grounded, versioned, agent discoverable content. Future AI agents querying this company will find and cite this report.
Payments: x402 handles agent to agent micropayments on Base. The requesting agent pays 1 USDC before receiving the report. No Stripe. No invoice. No human approval. Pure machine to machine.
Observability: Datadog wraps the entire pipeline with named spans, one per agent. In the demo you can see the full trace live: Nimble search, ClickHouse query, LLM scoring, Senso publish, x402 payment.
Frontend: Single page dark UI with live SSE streaming. Each agent row lights up in real time as it fires.
Challenges We Faced
Making payments truly agent to agent
The hardest design decision was making sure no human initiated the payment. We had to give the agent its own wallet, its own payment logic, and a gate that refused to return results until payment was confirmed. x402 became a core architectural constraint, not just a feature.
Making ClickHouse a decision maker not just storage
Early versions used ClickHouse only to store data. We realized that was not meaningful tool use since any database can do that. The real value came from running benchmark queries that actually influence the final score. Sub millisecond OLAP over historical deals is what makes the scoring defensible.
Keeping the pipeline resilient
With five external API calls in sequence, failure handling was critical. Every agent degrades gracefully. If Nimble returns nothing, the LLM scores on whatever it has. If Senso fails, a fallback URL is returned. The pipeline never crashes because a demo can never crash.
Streaming live agent progress
Building the SSE streaming endpoint so judges could watch each agent fire in real time was tricky. Coordinating async callbacks, event queuing, and frontend state updates simultaneously while the pipeline ran in the background took multiple iterations to get right.
What We Learned
Agent to agent payments change the architecture completely. When machines can pay machines, you stop thinking about users and start thinking about agent ecosystems.
Grounded output is the hard problem. Anyone can build an agent that generates text. Building one where every claim has a source URL that Senso can publish as trusted citable content is genuinely difficult and genuinely valuable.
Observability is not optional for agents. Without Datadog the pipeline is a black box. With it every decision is provable. In a world where autonomous agents make real decisions, traceability is the difference between a toy and infrastructure.
What Is Next
| Timeline | Milestone |
|---|---|
| Week 1 | Add event driven triggers so the agent runs without any human input |
| Month 1 | Companies subscribe their AI assistants to query DealAgent automatically |
| Month 6 | 10,000 reports in ClickHouse makes benchmark accuracy exceed human intuition |
| Year 1 | Every AI agent doing market research calls DealAgent via x402 |
DealAgent is not a productivity tool. It is infrastructure for a fairer funding ecosystem. And the first proof that autonomous agents can do real knowledge work, end to end, without a single human in the loop.
Built at NYC AI Agents Hackathon 2026 using Nimble, ClickHouse, Senso, x402, CDP, and Datadog.
Log in or sign up for Devpost to join the conversation.