Crossing Lanes. Part 1: AI Agents in Trading Operations

In one of our trading teams, a developer’s side project now answers the questions operations used to escalate. The same people close more incidents, and the team plans to keep its size as the workload grows.

Crossing Lanes. Part 1: AI Agents in Trading Operations
Figure 1. Alert handling before and after the internal bot.

A year ago, the roles on a trading team had clear boundaries. Operations specialists watched alerts and kept systems running. Traders came up with strategies. Algo developers turned those strategies into code. When someone needed something from another lane, they wrote a request and waited their turn.

Today those borders are much easier to cross. Here’s what that looks like on one of our teams at BHFT: AI agents have cut tasks that took days down to half an hour and let people take on work that used to belong to another function. Operations specialists run their own backtests. Traders build robots. Developers spend more of their day reviewing code than writing it.

The same pattern is showing up well beyond trading. In a study of its own workforce published in October 2026, Anthropic reports that coding-agent sessions among its non-technical employees grew from 2 a month in August 2025 to 90 a month in May 2026.

This "Crossing Lanes" series follows the change through the three groups on the team. This first part covers operations, where it started with one developer and a bot. Part 2 looks at developers, and Part 3 at traders.

A side project that took the questions

In operations, the change started with one developer who wanted to spend less time on support. He had built an automated system, and since he knew it best, every question about it came to him. To free himself up, he used Claude to write a bot that answers those questions. The bot, now called Optimus, runs on a self-hosted model.

What began as one person’s side project is now a shared effort. The developer and the operations team keep adding context to the bot so it can cover more and more of the questions that come up during a shift.

The alert workflow shows the difference. Before, an alert could go unhandled if the person on duty lacked the time, knowledge, or context to deal with it. The specialist would then have to find someone who understood the system, often a developer, and escalate. That was an expensive step, because it took developer time. Now the specialist asks the bot and gets a full breakdown of what to do.

Two effects are already clear: closing an alert now costs a fraction of what it used to, and the same number of people close noticeably more incidents.

Why automation finally pays off

The ideas behind today’s automation are not new. In early 2025, the team was already thinking about decision-support systems and ways to take manual steps out of operations. The obstacle was cost. Processes changed, people changed, and the markets the team traded changed, so an automation built for today’s setup could be out of date in a few months. The sensible choice was often to leave a chunk of operational work as it was.

By mid-2026, that math looked completely different. A typical example is a performance analysis across a large trading system. An operations specialist has to look at every asset in the set, understand how each one performs, and act on the result. In the past, this meant pulling data from several sources, writing SQL queries, and building a dashboard in a BI tool such as Tableau. The job could take several days. Today it takes a couple of prompts to Claude and 15 to 30 minutes.

Figure 2. The same operational task in early 2025 and in mid-2026.

The speed has a long history behind it. The team documented its systems, mapped its processes, stored data in databases, and thought hard about how to group and label it. None of that produced a breakthrough at the time. Today it is what lets an AI model work well. The better your data is organized and documented, the better your prompts and your AI-based automation work.

Industry research backs this up. Gartner predicts that through 2026, organizations will abandon 60% of AI projects that lack AI-ready data. Its survey found that 63% of organizations either do not have or are not sure they have the right data management practices for AI.

More work, same team

The second effect, more incidents closed by the same people, shapes the team’s staffing strategy. Operational work tends to expand, and the usual response is to keep hiring. The operations team regularly says it needs more people because there is not enough time. The team’s leadership has made a different choice: to keep operations at its current size and move as much context as possible into the models.

Understandably, some specialists feel uneasy about handing their work to an agent: it can feel like a threat to their job. The leadership’s message to them is that everyone’s tasks are changing, and a team that keeps working the old way will grow its headcount and lose its edge. Operational work exists everywhere. Once it is automated in one place, people move on to the next place where things are inefficient.

Where the freed-up time goes

The list of tasks an operations specialist handles has barely changed. The way the day splits between them has changed. Routine work and small development jobs now go to Claude, and the time that frees up goes to research, analysis, and finding problems early.

Research is the first destination. Specialists can now dig into why an alert fired, what sits behind it, and what else can be done about it. Some go further and step into another role’s territory. The most common example on the team is running a backtest. AI has made backtesting much simpler, so an operations specialist can pick up the tool, run a test, and draw conclusions from it. The first attempt rarely goes perfectly. The specialist tries, gets something wrong, asks questions, and does better the next time. For people who want to grow beyond operations, this is a low-risk way to try trading work and find out where they want to go next.

Anthropic sees a similar effect among its own staff. In a survey of its engineers and researchers, people reported becoming more “full-stack,” able to succeed at tasks beyond their usual expertise.

Conclusion

The operations story of this team comes down to three steps. Years of documentation made the data ready for AI. A developer’s side project turned that data into answers. The team’s leadership then chose to spend the saved time on research and growth and to keep its headcount stable.

The bot also changed who depends on whom. Operations used to wait for developer time. Now developers field fewer support questions, and their own work is shifting toward code review and shared knowledge. Stay tuned for Part 2, where we look at how the developer’s job is changing.

Sources

  1. Anthropic, “How employment is changing at Anthropic,” October 2026

  2. Anthropic, “How AI is transforming work at Anthropic,” December 2025

  3. Gartner, “Lack of AI-Ready Data Puts AI Projects at Risk,” February 2025