Inspiration

In finance, the first to know is the first to act. But LLMs are frozen at their training cutoff, and a single chatbot is too slow and too shallow to trace how one breaking story ripples through a supply chain to the stocks you own. A real trading desk doesn't work that way. It's a team of specialists: someone watching the wires, someone digging into what it means, someone making the call.

We wanted to build that desk as software: a swarm of agents that notices when the world changes, works out the knock-on effects for your Portfolio, and hands you a Thesis where every quote has been checked against its source. We call it First Strike.

What it does

First Strike watches your Holdings, reacts to breaking news, and delivers a cited Thesis for every Event that affects you. It has three roles:

  1. The Scout: Every 5 minutes it checks free news sources (Google News, RSS feeds and SEC filings) for each Watch Topic linked to your Holdings, such as "lithium" for TSLA and ALB. It posts every relevant headline as a Sighting on the OpenCrawl board.
  2. The Supervisor and Researcher: The Supervisor groups Sightings into Events, so 40 reports of one mine strike become one Event, not 40 alerts. It confirms each new Event with a single You.com search, then a Researcher investigates four questions: exposure, magnitude, timing and counter-evidence, using the You.com Research API. Every quoted passage is fetched and checked word-for-word against its source page before it counts as a Receipt.
  3. The Analyst: It writes one Thesis per Event, with a Stance (Bull, Bear or Neutral), an Action (Add, Hold, Trim, Exit or Hedge) and a confidence for each affected Holding. That means one lithium strike can come out Bear for an EV maker and Bull for a lithium producer, in the same Thesis. Claims without a verified Receipt are labelled as Inferences, and missing facts lower the confidence.

Every Thesis lands on a dashboard. High-confidence calls on your bigger Holdings also go to Discord. When a follow-up headline arrives, the swarm researches it and publishes a revised Thesis that lists what changed. And every call is graded against the real price move after 1, 5 and 20 trading days, so you can see whether the swarm is any good.

How we built it

First Strike is a Python app with a server-rendered dashboard, built quickly with Cursor and organised around a few deliberate design decisions, each recorded as an Architecture Decision Record.

  • You.com, the evidence layer: We use free feeds to find news and You.com to prove it. One You.com search confirms each new Event before any research money is spent. The Research API then supplies the facts and source URLs behind every claim. A shared query log means the same question is never paid for twice, even when two Researchers ask it at once, and a monthly budget the app enforces caps spend. Keeping detection free brought running costs down to tens of dollars a month.
  • OpenCrawl, the shared board: The Scout and Supervisor coordinate through OpenCrawl's message board. We designed the Researcher and Analyst to run as recorded Claude Code sessions, so the Analyst can read the Researcher's work word-for-word in OpenCrawl. For this build they run as built-in code inside the app, with each Researcher's Dossier stored locally and handed to the Analyst by ID.
  • A single Supervisor: The board can deliver the same message twice and has no locking. So the Supervisor is the only thing allowed to create Events, which rules out duplicate Events by design. It runs up to four Researchers in parallel.
  • Honest failure handling: Each stage of the pipeline runs separately and reports its own status. One broken feed or a failing Discord webhook shows up on the dashboard without stopping everything else.

Challenges we ran into

  • One story, forty headlines: Breaking news arrives as a flood of near-duplicate reports. Grouping them into a single Event, while still recognising a genuinely new development as an update to that Event, took several iterations.
  • Cost: Polling a paid news API for every topic every few minutes would have cost hundreds of dollars a month. Moving detection to free sources, and using You.com only where verification matters, solved it.
  • Citations that actually prove something: A link is easy to fabricate and easy to misread. We ended up fetching every source page and checking that the quoted passage really appears on it, and labelling everything else as an Inference.
  • Revisions: Our first version kept reusing old research when a follow-up headline arrived. We had to version Dossiers so a revised Thesis is built on the new evidence.

Accomplishments that we're proud of

  • Verified, not just cited: Every quote in a Thesis has been fetched and matched word-for-word against its source. Anything we can't verify is clearly labelled as the Analyst's Inference.
  • One Event, opposite calls: The swarm reasons per Holding, so it can tell you a story hurts one stock and helps another, instead of flattening it into one sentiment score.
  • It keeps score: A built-in Track Record grades every Stance against the actual price move after 1, 5 and 20 trading days.
  • It runs on its own: No prompting needed. It watches, groups, researches and publishes by itself, and tells you plainly when a stage is failing.

What we learned

We learned that the hard part of a multi-agent system isn't the reasoning, it's the plumbing: how work is handed off, how duplicates are avoided, and how context survives between agents. We learned to separate finding news from proving it, which made the system both cheaper and more trustworthy. And we learned that a citation is only worth something if you check it. Cursor let us change the architecture several times during a 10-hour sprint without losing momentum.

What's next for First Strike

  1. Recorded agent sessions: Run the Researcher and Analyst as recorded Claude Code sessions, so OpenCrawl holds every agent's full working history, word-for-word.
  2. An LLM Analyst: Replace the built-in rules with a model that reasons over the Dossier, held to the same Receipt and Inference rules.
  3. A Devil's Advocate agent: A fourth agent whose only job is to argue against the Analyst using historical market data before an alert goes out.
  4. Paper trading: Connect a broker such as Alpaca for simulated trades based on each Thesis, so the Track Record can measure returns, not just direction. First Strike will not place real trades.

Built With

Share this project:

Updates

Submission history