AI Hackathon
← DEPARTURES
US0105AI AgentsBoarding

Google - Agent Harness on Physical-Engineering's Last Exam

Google LLC · on Kaggle

Cash prize: $400,000
Free credits: Credits: not stated
Who can enter: Solo OK Country rules checked
Code rights: Not checked yet
GoogleGoogle Foundation Models
Submissions closeJan 5

Entrants build autonomous agents that work through hard structural, civil, electrical and mechanical engineering problems.

Proceed to gate Opens kaggle.com. You register on the organizer's site.

Imported from Kaggle and approved by our team. We haven't checked the rules by hand. Confirm the rules on the event page before you enter.

One email a week

New events that match your topics, every Monday. Unsubscribe anytime.

Flight details

Submissions close:
2027-01-05 23:59 UTC
Event ends:
2027-01-05 23:59 UTC
Cash prize:
$400,000
Who can enter:
Open worldwide to individual entrants and teams up to 5, excluding US-sanctioned jurisdictions.
Page language:
English

Prizes listed: Total cash prizes: USD 400,000 ($400,000 total cash prizes: $250,000 for 1st place, $100,000 for 2nd place, and $50,000 for 3rd place.); Credits (unknown): amount not stated (Not stated on the event page.)

In short

  • Final submissions are due January 5, 2027 at 11:59 PM UTC, and entrants should confirm the dates on the organizer page.
  • The total cash prize pool is USD 400,000, split as USD 250,000 for first, USD 100,000 for second and USD 50,000 for third.
  • Final submissions are scored against 56 hidden private test tasks, and winners must release code under an open source license.

Project ideas

Multi-Discipline Design Agent

A harness that splits a long structural or mechanical design brief into checked sub-tasks before it commits to an answer.

Architecture: Parse the brief into discipline-tagged sub-questions → Retrieve the formulas and code clauses each sub-question needs → Run Python solvers and verify units at every step → Assemble a cited answer and flag low-confidence sections

Stack: Google Foundation Models, Python, SymPy, pytest

Code Clause Retrieval Agent

A retrieval agent that grounds each engineering answer in quoted standards text rather than relying on model memory.

Architecture: Chunk standards and handbook excerpts into a searchable index → Agent writes search queries from the parsed problem statement → Rerank passages and attach a citation to each calculation → Reject any answer whose passage support falls below threshold

Stack: Google Foundation Models, Python, FAISS, Sentence Transformers

Hidden-Task Robustness Harness

A harness tuned to survive unseen private tasks by stress-testing itself on self-generated engineering variants.

Architecture: Generate perturbed variants of every solved practice task → Run the agent on variants and log failure categories → Route the weakest categories to extra tools and prompts → Freeze the best configuration and rerun the regression suite

Stack: Google Foundation Models, Python, Docker, Weights & Biases

FAQ

Can I enter on my own, or do I need a team?
Solo participation is allowed, and teams may have up to five members. The event page states the maximum team size is five participants.
Who is allowed to enter this hackathon?
The page says the event is open worldwide to individuals and teams of up to five, excluding US-sanctioned jurisdictions. Eligibility is not confirmed, so check the official rules before registering.
Who owns the code I submit?
Winners must deliver final submission code under an open source license, Apache 2.0 for harness code and CC BY 4.0 for writeups. Ownership terms are not confirmed, so read the official rules.

Sources

Open the event page →

More open AI hackathons