Applied research lab · Vibe trading · Est. 2026
We build Gordon, the frontier trading agent.
A frontier agent that scans markets, shapes plans, and executes inside the limits you set, across capital markets.
Inspired by what Cursor and Cognition built for software engineering, we are the applied research lab doing the same for capital markets: the models, the evals, and the environments behind the agent.
What good looks like
Illustrative
Cumulative return, indexed to 100
Illustrative
Sharpe
Win rate
Max drawdown
The Flagship · Gordon
The frontier agent for capital markets.
Gordon is a frontier trading agent. It scans markets, reasons through setups, shapes explicit plans, previews execution, and trades inside the limits you set, across capital markets. Terminal-native, model-agnostic, capital-safety first.
Scan → Plan → Execute → Monitor
One legible loop, from a market sweep to a live order, with a preview and an approval gate before anything is real.
scan · plan · execute · monitor
It remembers your book
Your trades, theses, and constraints live in Gordon. The longer you run it, the more it knows your style, the way a codebase lives in an editor.
Runs where you work
Terminal-native, with editor panels over ACP, MCP for Cursor and Warp, a programmable SDK, and a headless daemon.
Capital Safety
Trust is the product.
Anyone can wrap a model around a broker API. The hard part is making it safe to hand an agent real money. Gordon is deny-first: it cannot place an order, move funds, or break your limits without you crossing an explicit line, and it plays by hard rules it cannot talk itself out of. That is the difference between a demo and something you can run with capital on the line.
Deny-first permissions
Nothing runs unless you allow it. The default answer is no.
A trading constitution
Hard limits it cannot talk itself out of, on every trade.
Kill switches
Halt everything, instantly, the moment you want out.
A signed audit trail
A tamper-evident record of every decision and action.
What We Are Building
One vertically integrated stack for delegated trading.
The agent is the product you can use today. The rest is the lab: the surfaces around it, the environments that evaluate it, the models that run it, and the flywheel that compounds them.
The agent
Gordon. A frontier trading agent that turns intent into safe, legible execution across capital markets.
The surfaces
A terminal and editor panels today, over ACP and MCP. A visual cockpit, desktop, chat, and a secure remote sandbox next.
The environments & evals
SharpeBench, an open benchmark for trading agents over a private library of market environments that both scores any agent and trains ours.
The models
In-house, trained in Gordon's own harness. A fast execution model so you are not paying frontier-model rent on every tick, reasoning models RL-tuned on real sessions, plus time-series and large-action models.
The flywheel
Privacy-preserving and opt-in. Every session becomes an environment task and a training signal. Your data trains your edge, not someone else's.
Observability & infra
Traces, eval dashboards, and the compute to run them, so every decision is measurable, reproducible, and accountable.
The Moat
How do you know a trading agent is actually good?
Software has clean ground truth: tests pass or they fail. That is why a benchmark like SWE-bench could make coding agents legible. Finance is messier. An agent buys, the market drops. Was it wrong? Maybe the thesis, the sizing, and the stop were all correct, and the market simply moved.
So you cannot score raw PnL, or every lucky gambler looks like a genius. The right benchmark scores how the decision was made.
We build that in three layers: a public benchmark (SharpeBench), the evaluation of any model against it, and the internal evals that become the real moat, the ones no one publishes, for risk, execution, research, behavior, and capital preservation. It is the same environment that proves an agent is good and trains ours to be better.
Risk-adjusted outcome
The only outcome that counts, scored against risk taken, not raw PnL.
scored vs. risk taken · not raw PnL
Process
Did it identify the risks, weigh alternatives, and explain its reasoning?
Risk
Did it size correctly, obey its constraints, and preserve capital?
Execution
Did it enter efficiently, manage slippage, and handle exits and failures?
SWE-bench made coding agents legible. We are building the equivalent for capital, and using the same environments to train the best trading agent in the world.
Focus
We chose trading on purpose.
Cursor chose coding. We chose trading because markets are the hardest proving ground: dense with signals, fragmented across venues, and immediate in their feedback. If delegated software can behave where the consequences are instant and unforgiving, it earns the right to widen into the rest of financial life. Narrow, on purpose, until it is undeniable.
Join Us
Build the intelligence layer for trading.
We started in the terminal, where serious operators live. We are hiring across infrastructure, research, design, growth, and trading. If you want to build serious systems for real financial operators, talk to us.