The Playbook Already Exists
Over the last three years, software engineering was rebuilt around agents. Cursor and Cognition did not win by wrapping a chat box around an editor. They built a product operators actually live in, then a research lab underneath it: in-house models trained on real usage, evaluation systems that defined what good even means, and a data flywheel that compounded with every session. The benchmark became a category. The model became a moat. The product became the training environment.
Capital markets have not had that moment yet. The tools are still built for humans clicking through terminals, spreadsheets, and dashboards. Coding-agent evaluation has moved to end-to-end, real-environment tasks, and what gets scored is the whole agent system, the model plus its tools, not a model in isolation, but markets are still waiting for the equivalent of what coding got. General Liquidity exists to build it.
The scale of that shift is easy to forget. When the first version of SWE-bench appeared in 2023, the best system resolved about two percent of real GitHub issues. Once the benchmark was hardened into a curated, human-verified set, the frontier climbed from roughly half of those issues to three-quarters in under two years. Trading has had no such curve, because it has had no such benchmark.
Intelligence Is Not Enough
The common mistake is assuming a capable model plus market access equals a capable trader. It does not. A model that speaks fluently about a setup is not therefore able to size a position, respect a limit, recover from a failed order, or preserve capital through a drawdown. Agent reliability benchmarks score success across many runs, not on average, and show how fast run-to-run inconsistency breaks the illusion that capability alone is enough. In most software, being slightly wrong is survivable. In markets, being nearly right can still be the same as being wrong.
The arithmetic is unforgiving. A single trade is not one decision but a chain of them: read the state, size the position, route the order, confirm the fill, reconcile the result. The trade only succeeds if every link does. If each step is reliable with probability p and the chain is k steps long, the chance the whole thing lands clean is p to the power of k, and it decays faster than intuition expects.
So the hard part is not the model and not the venue. It is making the full loop trustworthy enough to run with real money, and each piece of that loop exists to bound a specific failure. Reasoning with explicit uncertainty stops false confidence from sizing a position. Hard permissions and an approval step stop an order that was never sanctioned. Durable memory and replay stop the system from forgetting its own state between ticks. Disciplined execution and clear failure modes stop a single bad fill from compounding into a cascade. None of these is a model capability. Each is a property of the harness around the model, and the harness is the work.
Why Trading First
We start in trading because it is the hardest place to get this right, which makes it the right place to prove it. Trading is live, demanding, and unforgiving, and it gives immediate, honest feedback. It does not let us hide behind a good demo. If a system can be trusted with capital here, it earns the right to widen across the rest of capital markets later.
This is also where the shortcuts fail most visibly. Pure language systems are strong at explanation and weak at numerical reasoning, time-sensitive context, and memory that survives a session. New rails settle faster without solving trust. Chat makes software feel simpler without making the workflow safer. When the stakes are low those gaps are easy to ignore. In markets they are immediate.
An Applied Product and Research Lab
General Liquidity is an applied product and research lab, and the two halves are not separable. A pure research org produces elegant ideas with no operator pressure. A pure product org ships shallow features on whatever model is strongest this quarter. We want the product to pressure-test the research and the research to keep the product from collapsing into demos. The strongest financial-agent systems already separate a dedicated risk and verification role from the decision-makers, which is exactly the posture markets demand.
Our belief is that the moat is not access to one model or one venue. It is process engineering: understanding how serious market work actually gets done and encoding it into a system trustworthy enough to use with capital on the line. Capital safety is not a feature we add at the end. It is the architecture. Gordon is deny-first, plays by hard limits it cannot talk itself out of, and writes a signed record of everything it does.
Gordon Comes First, and Today It Is the Product
Gordon is the first and, today, the only product. It is a frontier trading agent that runs in your terminal: scan, analyze, plan, preview, approve, execute, monitor, reconcile. It is terminal-first because the earliest serious users care about clarity, speed, auditability, and control, not ornament. We are not pretending Gordon is a universal financial operating system. It is a sharp wedge with immediate feedback, and that is the point.
We are honest about what exists. Gordon ships. The rest of what follows is our research direction, not a live product, and we would rather name it plainly than imply it is already built.
What We Are Building Behind It
The same loop that lets you run Gordon is the loop that lets us improve it. Every session, with consent, becomes a trace and a training signal. That points at a few pieces of research we intend to build, in order, only as the wedge earns them.
The first is models. We expect to train in-house models inside Gordon’s own harness: a fast execution model so the hot loop does not pay frontier-model rent on every tick, and reasoning models tuned on real sessions. The second is evaluation. Software engineering keeps shipping serious, evolving agent benchmarks; trading still has no standardized equivalent. We intend to build SharpeBench, an open way to ask whether a trading agent is actually good, scored on process, risk, execution, and risk-adjusted outcome rather than raw PnL, alongside the private internal evals that become the real moat. The third is environments: historical and synthetic markets, including the crises, that both grade an agent and train one. The fourth is the flywheel that ties them together.
None of that is shipped. It is the work, sequenced behind a product that has to be real first.
Earning the Right to Widen
The order matters. The broad claim comes last, not first. Inspired by what Cursor and Cognition built for software engineering, we are doing the same for capital markets: a product operators live in, the models and evals beneath it, and a flywheel that compounds. We build where the stakes are highest, make it trustworthy there, and expand outward with a right that has been earned rather than narrated. Gordon is where that work starts.