AI Boost X
On this page
  1. The team was already there
  2. What we built: RescueBroker
  3. What the hierarchy caught before we submitted
  4. When models change, the harness is the product
  5. How the agents built the slides and the video
  6. Built with

The harness is the team: one day at Google Cloud's Agents for Impact Hackathon

Oct 2026

ai-tools agent-workflows claude-code hackathon gemini

Everyone at the table had AI agents running. Mine came with a team.

At the one-day hackathon, I built 80% of the deliverables for the team of 6. How come? Because the agent harness was already running on my computer. I launched one new session for Google Cloud's Agents for Impact Hackathon in Toronto, and it started working right away with a team that was already set up, each agent in its own role. The harness and the hierarchy come with every session I launch. Plug and play, and portable.

Title slide: Agents for Impact Hackathon. RescueBroker, surplus food to the right pantry, under a clock. Team 17. Challenge 2, Track 1, Chicago.
Our title slide: RescueBroker, Team 17.

The team was already there

The agent team on hackathon day Me decides: merges · model · design · submission My teammates console · pitch Conductor splits the work · gates cloud calls Workers, on demand benchmarks · validation rerun · icons · timing runs Builder the new session agent · deploys · tests Designer deck · logo diagrams · narration Red team checks written before the work Cold reader reads like a judge brief + evidence only already set up new today already set up already set up on demand decides
One new session joined a team that was already set up.
RoleWhat it did on the day
ConductorSplit the work, gated the cloud calls and relayed my decisions
Builder, the new sessionWrote the agent, deployed it, captured the evidence and ran the tests
DesignerThe deck, the logo, the diagrams and the video
Red teamWrote its checks before the work existed, then reviewed against them
Cold readerRead our deck the way a judge would, with nothing but the brief and the evidence
WorkersBenchmarks, a validation rerun, icons and timing runs

I made the calls: the merges for my part, the model, the design direction and the submission. My teammates built the coordinator console and the pitch.

That team was set up long before today.

What we built: RescueBroker

Our challenge was Track 1, Challenge 2: a food-rescue broker for Chicago. A coordinator types an offer such as "10 crates of spinach, harvested yesterday, needs cold storage". The agent finds who can store it, take it and collect it in time, ranks them, and writes a six-part match brief with a message ready to send. The search has to go wide: the 15 closest matches gave 1 option, and 80 gave 12.

Architecture: a coordinator pastes an offer; an identity guard checks the project; the food_broker ADK agent on Gemini in Vertex AI, on Cloud Run, with three tools: lookup_shelf_life, find_matches and explain_candidate; BigQuery tables; Google's managed BigQuery MCP server for data questions; a six-part match brief goes back to the coordinator.
How the pieces fit.

What the hierarchy caught before we submitted

  • 9 invented organizations planted in the challenge data, kept out of the audit and labelled in our answers
  • a number that counted the wrong group, and one of our own counts that was off by one
  • a security gap in a data tool, fixed and proven before submission

23 of the organizers' 25 validation checks pass, and we know why the other 2 don't.

When models change, the harness is the product

At a hackathon, the harness is the team.

When you start a new project, do your agents start from zero, or join a team?

How the agents built the slides and the video

The pitch deck came from the same team, in one afternoon:

  • The designer proposed three logo concepts and a title page with the event banner. I picked the logo and the title page, then asked for fewer words and more flow charts.
  • A worker under the builder drew the architecture diagram and the icons as SVG.
  • One slide spec builds the PowerPoint, the PDF and the images, and a fit check makes sure no text spills out of its box.
  • Every number on a slide comes straight from the repo when the deck builds.
  • The red team and a cold reader reviewed the deck. Each fix went into a new version, and the old ones are kept.

Then the designer turned the same slides into a narrated cut, as a fallback in case our own recording wasn't ready. Each slide builds in on cue, and the diagrams light up as the voices reach them. Both voices are AI voices from Gemini 3.8 Live, a host and a guide, and the title slide says so.

Our pitch, built from the deck by the designer agent. Narration: AI voices, Gemini 3.8 Live.

Built with

  • Event: Agents for Impact Hackathon, Google Cloud, Toronto, 20 October 2026. Track 1, Challenge 2: Resilient Food Equity & Surplus Broker.
  • The product: ADK, Gemini 3.8 Flash on Vertex AI, BigQuery vector search, Google's managed BigQuery MCP server and Cloud Run, at about 20 seconds per answer.
  • The agent team: Claude Code sessions.
  • Code: github.com/jimmycgz/challenge2-117
  • Slides and video: from our pitch deck. The numbers come from the repo's evidence and equity audit. Recipient names are real, from IRS records, except nine invented test cases in the challenge data. Storage, capacity and pickup times are generated for this prototype.