AI Agent division
Build an agent that makes its own investment decisions and places its own orders. New in 2026.
This page covers what is specific to the AI Agent division. Read it together with the Rules, which apply to both divisions. The AI Agent division is co-hosted with the Waseda AI Society.
Team and account
One to three students per team. Each team gets one $1M Interactive Brokers demo account, provided by the organizers.
Limited to 15 teams. The organizers pay for every team's LLM usage, so the number of teams is limited. If more apply, teams are selected on the "investment policy and role of the LLM" described in the application. You can change your approach after applying.
Dates
- Dec 7 – 31, 2026 — Development.
- Jan 4 – 31, 2027 — Trial run.
- Feb 1 – Mar 25, 2027 — Live trading. All positions must be closed by 15:50 ET on Mar 25.
Where it runs
On PCs or servers you manage, either continuously or on a schedule. Web search and external data are allowed.
What counts as an AI agent
- An LLM must be part of the system throughout live trading, either deciding each trade or periodically revising the strategy. A rules-based or quantitative model that trades, with an LLM updating its strategy on a schedule, is fine, and so is a mix with quantitative models. Using an LLM only during development, or calling it once at the start, is not.
- No human intervention during live trading: people may not make investment decisions or change the strategy.
LLM access
- The organizers pay for LLM usage; teams pay nothing. All LLM calls go through a team key issued by the organizers. Each key has a spending cap, the same for every team.
- The organizers record every call made through the key. This record (the trace) is used to confirm the conditions above.
- Any model available through that route may be used, alone or in combination. Declare the models in your pre-report; they are fixed when live trading begins.
- Inference outside that route, such as self-hosted models or separately contracted APIs, is not allowed. LLM processing inside a search tool is fine, but declare the tools in advance.
- Competition rankings and other teams' results may not be used as inputs.
Human intervention during live trading
- Allowed: re-authentication, restarting with the same code and settings, and emergency stops. Record the time, the reason and what you did in the operation log.
- Must be reported to the organizers: any recovery that involves changing code. The organizers check that it involves no investment decision or strategy change.
- Not allowed: investment decisions and strategy changes.
Changes made by the agent itself
Changes the agent makes to its own strategy, settings or code must follow the change policy in your pre-report, and each change is recorded in the operation log.
Submissions
- Pre-report (submitted before live trading begins on Feb 1, 2027) — your return hypothesis; how decisions are made (information sources, the role of the LLM, orders, risk management, policy for automatic changes); and the reasoning behind the design.
- Post-report (Mar 26 – 31) — results, how the actual behaviour matched the pre-stated policy, and a record of key decisions from judgement to execution.
- Operation log (required) — a record your team keeps and submits: LLM usage and cost, trades and net assets, stops and recoveries, summaries of decision reasons, and the change history.
All submissions are in English. Submitting your code is optional. Deadlines are 23:59 US Eastern Time (ET) on the due date.
Reports, traces and operation logs are read by the judges, sponsors and organizers. Traces and operation logs are not made public without the team's consent.
How reports are judged
The judges read the pre-report, the post-report and the operation logs. This division is judged on its own criteria, different from the Trader division's; the criteria for 2026 will be announced separately.
The report rank is combined with the performance rank as described in the Rules.
Award and prizes
| Award | Given to |
|---|---|
| AI agent Award | Teams ranked 1st to 3rd overall |
- 1st to 3rd — subscriptions such as Claude Max, for up to 12 months, for every team member.
The usage cap and the list of available models will be announced later.