Agents on the desk
What our agents actually do day to day, and the line they don't cross. They carry the surface area; people keep the wheel on anything that moves capital.
Narrated in the writer’s own voice · 2:52
Contents4 sections
The same tooling we use to build the desk also helps run it. That is the point, and it isn't about novelty: the people building the system and the agents operating it share one loop. Nothing gets thrown over a wall to a separate operations layer. Research feeds tooling, tooling feeds the desk, and the desk feeds back into research, all through the same set of agents.
What they do
Day to day, the agents do bounded, checkable work, and they do it well:
- run backtests and summarize the results;
- watch logs and metrics, and flag when something drifts;
- draft research notes and check each other's;
- keep the recurring chores moving;
- catch the boring failures early — the stale feed, the job that didn't fire, the number that stopped updating.
This is unglamorous work, and it is exactly the work that quietly eats a small desk alive when nobody is watching it.
The line they don't cross
There is a hard line, and it doesn't move. The agents propose, prepare and surface. People stay on the decisions that move capital, and on anything that touches risk limits or the live order path. No agent holds a key to a venue; the most an agent can do is lay a prepared order in front of a person, with its reasons attached.
An agent that can be wrong should never be the last checkpoint before the market.
We treat it as a rule, not a preference. Agents assemble the context and lay out the options; a person makes the call and takes the action. In the strictest mode that is two people: an order goes out only when both have approved it.
Leverage, not autopilot
For a desk of two, this is leverage. The agents expand how much ground two people can cover: more research explored, more monitored, fewer things falling through the cracks between one session and the next. Decisions arrive on a steady ten-minute cadence, already prepared, rather than as a stream of interruptions. What the agents do not do is take over the calls that need judgment. We aren't handing off decisions; we are getting them surfaced better prepared, with less noise in the way.
The shape we are after
A desk where people are amplified, not replaced. The agents carry the surface area. The people keep the wheel, stay close to the risk, and own every decision that reaches the market. Short feedback loop, clean incentives, and a clear line about who is responsible for what.