September 24, 2026
Share on
The NXT cohort, members of Citadel’s development program for engineers who are 2 to 5 years into their careers, gathered for their recent two-day offsite at the 1 Hotel Brooklyn Bridge in New York. Founded in 2018, the program has grown to include more than 200 alumni and active members—roughly a quarter of Citadel’s engineering workforce. The offsite, NXT’s annual capstone event, is designed to bring participants “off the desk” to build internal networks, sharpen technical skills and interact with senior leadership. This year’s offsite featured the program’s first AI Hackathon.
Each team was tasked with building an AI agent that would be commercially impactful to the firm. Attendees were assigned to cross-functional teams to surface common challenges in their work, then chose which ones their agent would address. This was an opportunity to take their Claude Code skills to the next level.
Senior engineers served as the teams’ coaches during a roughly 7-hour build period. The following morning, all 23 teams pitched in two-minute slots to Citadel Chief Technology Officer Andrew Janian. Only four teams advanced to present a full demonstration of their agent to the Hackathon panel.
Leading into the Hackathon, Andrew took the stage with Chance of the Citadel X team to walk through how he uses Claude Code in his own life. One example was a recipe app he built with Claude that helps him cook meals by working backward from the time he wants to eat. It plots every step for him, down to when to start boiling the water for mashed potatoes.
“I was not able to do it before Claude,” Andrew said, noting how much AI has changed the cost-to-value math on engineering projects. “I guess it wasn’t worth the time.” Because of AI, the cost of standing up a prototype, he noted, has effectively fallen to zero, making it possible to create the cohort’s working agents for 23 teams in a single afternoon.
Andrew demonstrated how he writes guided instructions, specifying the API shape, the interface and the payload. The same discipline applies to product code at Citadel. “My prompts don’t look like vibe-coding prompts,” he said. “I’m instructing it how I want it done, where I want the endpoint to be, how I want the structure to be.” As AI drives down the cost of writing code, the more valuable work moves upstream into clearly specifying what a system should do. Teams took those tips into the build period.
Attendees voted Ticker Talk the winning project. It’s a prototype that turns the daily flood of Slack messages, emails and market news into a fully customizable audio brief.
The setup occurs in a simple chat interface. You could tell it you are an analyst who wants a 10-minute brief at 6:45 a.m. covering the latest updates on energy and rates, or an operations manager who wants a 5-minute recap of overnight emails, messages and tickets. The agent produces a ready-to-play audio file. The team demonstrated it live with different role profiles.
The group also shared ideas on how it could keep evolving. “Imagine you get a phone call from the agent in the morning,” Calin (Post-Trade Engineering | ’25 NXT) said. “Then, when you hear something interesting, you can start to have a back-and-forth conversation based on what you need in the moment.”
Finding the right dataset often requires tracking down a colleague who has used it before. Kevin (Core Engineering | ’24 NXT) explained, “Sometimes we get messages like: ‘Do we have data about this?’ or ‘Who can point me to somebody who knows about this dataset?’”
Team Data Cat, a Hackathon finalist, presented an agent that aims to replace the ask-around process. Every new dataset is automatically described and tagged on arrival, and anyone can ask plain-English questions and get answers through a chat interface.
Building it required minimal manual coding. “We just set up the scaffolding and defined the tools and instructions,” Wistan (Equities Engineering | ’24 NXT) said. He walked through how they described what they wanted in words, and the AI wrote most of the code. When implementation is cheap, the design surface becomes the tools an agent can reach for and the instructions that govern how it uses them.
A finalist group from Post-Trade Engineering presented its Treasury Intelligence prototype, an AI agent that converts the roughly 400 routine reports that Treasury emails over two weeks into a structured, searchable dashboard with a built-in chatbot. The interface keeps raw emails visible so that users can trace any claim back to its source. That’s a deliberate design choice, as an agent that summarizes and interprets is only as useful as a user’s ability to check what it produced.
Charlie (Fixed Income & Macro | ’25 NXT) noted that the same pattern would work anywhere a team scrolls through recurring email reports to get the data they need.
The Second Brain team tackled a request from a stakeholder to “upload an investor’s brain into our AI tools.” They built a set of agents that maintain a living notebook on a colleague’s behalf, reading new filings, earnings calls and meeting notes as they arrive and automatically updating the notebook.
They built the prototype so that fundamental equities investors can track their investment theses. Second Brain noted that the approach works for any role that maintains an evolving body of knowledge, i.e., onboarding documentation that updates as policies change, a team leader’s playbook or institutional knowledge that lives in only one person’s mind. “It evolves and changes with information that comes in,” Kelly (Fixed Income & Macro | ’25 NXT) said.
Across the four finalists, each team spent less energy writing code and more on the questions AI makes newly important, like what an agent should be allowed to do, how much latitude to give it, which systems to expose as tools it can reach and how a user can trust the result. Those are the design problems that define building with AI at Citadel.
