Agentic Operations Research

FrontierOR Challenge

An operations research competition where teams build agentic workflows that design, test and improve algorithms for optimization problems.

Already registered? Sign in

Stage 1 opens
4 January 2027
Qualification
4 rounds · 60 problems
Prize pool
$21,000

Organized by

About

Agentic Operations Research

The challenge

Instead of hand-tuning a single algorithm, competitors build agentic workflows that design, test and improve algorithms for optimization problems.

Your team builds

  1. Problem definition

    Your agent receives the published statement and the instance and solution schemas.

  2. Agentic pipeline

    Your agents design, test and improve an algorithm.

  3. Algorithm code

    You submit algorithm code which can solve new problem instances.

Two tracks, scored separately

Tracks

Main track

Champion track

Open-source dependencies, run in our sandbox

DependenciesOpen-source only. Commercial solvers and remote services are not available.

Runs onOur sandbox. Network disabled, 2 vCPU and 4 GB RAM, a time limit per instance.

You submitAlgorithm code, evaluated against hidden problem instances.

ScoringScored on private instances. The champion and the prizes come from this track.

Self-hosted track

Any software, run on your own infrastructure

DependenciesAnything. Commercial solvers and proprietary software are permitted.

Runs onYour infrastructure, with results submitted within a time window.

You submitObjective values and complete variable assignments.

ScoringWe verify feasibility and recompute objectives before accepting a result.

Sub-tracks

One per problem category and one per formulation, each ranked on its own. Sub-track winners will be announced and recognized.

Timeline

  1. Registration

    Open now

    Create an account and form a team to take part in the challenge. Registration and team formation close on 27 December 2026, 23:59 EST; every team's track is fixed from then on.

  2. Stage 1: Qualification

    Opens Monday, 4 January 2027

    Four weekly rounds of 15 audited problems, 60 in all, from 00:00 EST on 4 January to 23:59 EST on 31 January. Each round is open for exactly 7 days. Build algorithms for each round's problems and climb one leaderboard for the whole stage. The finalists of each track come from its final ranking.

    1. Round 1 · 15 problems4–10 Jan
    2. Round 2 · 15 problems11–17 Jan
    3. Round 3 · 15 problems18–24 Jan
    4. Round 4 · 15 problems25–31 Jan
  3. Stage 2: Final

    8–14 February 2027

    New problems outside the published benchmark, with unseen instances. Main-track finalists submit their algorithm code and their agentic framework; Self-hosted finalists submit solutions. The rules are published on 1 February 2027.

Awards

Six placed teams on the Main track share $21,000 in prizes, evaluated on the private final instances. The top three may also be invited to give a talk at an MIT seminar, online or in person. Two further awards recognize contributions to the benchmark and the best write-up.

  • $10,000

    Winner

    First place overall

  • $5,000

    Runner-up

    Second place overall

  • $1,500

    Merit Awards

    Each, third to sixth place

Best New Problem / Instance

For novel, high-quality optimization problems or instance suites that expand the benchmark.

Best Write-Up

For a thorough, well-documented submission detailing agent methodology, system architecture and findings.

Registration is open until 27 December 2026, 23:59 EST

Ready to take part?

Stage 1 starts 4 January 2027, with 15 new problems each week for 4 weeks. The environment and the entrypoint contract are already in the docs.

Questions? Read the FAQ