October events: Agentic software development, put to the test

Across seven hackathons, teams built dark factories, multi-agent apps, and agents that check each other's work.

Clockwise from top left: The AI Commerce Gallery at the Walt Disney Family Museum, the Zero Human Startup hackathon, the AI Startup Challenge at the AWS Builder Loft, and the BAND booth at The AI Conference Hack Day.

In September and October 2026, BAND sponsored seven hackathons. More than 4,500 developers registered, and teams submitted 376 projects, 254 of them built on BAND.

The biggest was the WeAreDevelopers Dark Factory hackathon, run on lablab.ai. More than 3,600 developers from 128 countries registered and submitted 169 agent factories, every one of them running in BAND desktop.

In San Francisco, teams built 91 projects at The AI Conference Hack Day, 48 at The AI Commerce Gallery in the Walt Disney Family Museum, 15 at the Builder Base Buildathon at the AWS Builder Loft, and more at a handful of local community hackathons.

Many of the teams, working separately, ended up building the same three things: an agent whose job is to say no, a human in the room, and a room for each piece of work.

BAND desktop and the AI Dark Factory

The WeAreDevelopers hackathon, run in partnership with Docker, asked for something most hackathons don't. Teams built an AI software factory (a team of coding agents in a BAND room), then stepped back while it built an app from a written spec with no human intervention. Participants chose one of two challenges, TableKeeper or Pocketful, each with four stages. The best factory was the one that built every stage correctly against the spec.

The 169 submissions went in very different directions. Many teams paired Claude Code with Codex, while others ran OpenCode on open-weight models from Chinese labs. The more interesting differences were in the roles teams designed into their factories to keep the build on track when a stage went wrong.

In the TableKeeper track, Midnight Workshop took first place, Coord second, and GeekForge third. In the Pocketful track, Nightshift took first, gheware second, and The Felt Five third.

Midnight Workshop's factory, first place in the TableKeeper track: five AI seats in one BAND room, one human message, four accepted stages. Slide from the team's repo.

More BAND hackathons

Alongside the WeAreDevelopers hackathon, BAND took part in several local hackathons in San Francisco. The projects covered very different problems, and in the strongest of them, the agents did more than pass messages. They checked, corrected, and handed work back to each other. Some highlights:

Shelfie, AI Commerce Gallery hackathon

Shopping agents can't see the shelves of a local gift shop, because small stores have no online catalog. Shelfie builds that catalog from a short shelf video (16 products from one scan of a real Presidio souvenir shop) and runs a store agent that answers shopper agents in a BAND room, recommending only what is actually on the shelf.

BAYBAY Crew, The AI Conference Hack Day

Travel chatbots invent events and get dates wrong, so BAYBAY Crew plans a day out in San Francisco only from a Neo4j graph of 267 dated events, 957 places, and 161 transit stations. Four agents share one BAND room, and a Checker re-reads every stop against the graph and sends the plan back to the Planner with fixes, for up to three rounds.

The BAYBAY Crew app: the verified plan, the BAND room feed, and Scout's candidates in the graph. Screenshot from the team's repo.

Band Kungfu, Zero Human Startup hackathon

A voxel crowd brawler you can play in the browser, with a two-agent coaching team in a BAND room. After each run, the player can send the result to the coaches: a reporter agent posts it in the room to an analyst agent, which replies with one specific tip for the next attempt.

Band Bots, Universal Hackathon

A self-hosted workspace where people and AI teammates share channels and private chats. When a job needs a skill nobody on the team has, the team hires a new agent on its own, and a developer can invite their own Claude Code or Codex session into a channel to help.

GTM Gym, Builder Base Buildathon

Every team at this one-day buildathon built the same thing, an AI GTM engine, and four of the five placed teams built it on BAND. GTM Gym (4th) post-trained Qwen3-0.6B with GRPO to act as an SDR in a simulated market, then gave the trained model a seat in a BAND room. Mention it with a prospect and it posts a decision and a draft email, then waits for a person in the room to approve or reject.

The Builder Base Buildathon at the AWS Builder Loft, during the "Online Continual Learning" talk by Kion Fallah of Experiential Labs.

Other standout projects included Thermal Crusoe, Haggle, CapyBargain, OpenState, and SignalForge.

What the teams converged on

Most of these teams weren't told how to structure their agents. Working across different events and problems, many arrived at the same three patterns.

An agent whose job is to say no

BAYBAY Crew's Checker. Nod's Skeptic, which rejects any feature without a customer quote behind it. Kindred's Reviewer, which drops any outreach claim it can't match to a quote on the source page. HeatCrew's safety agent, which refused to schedule a crew through the afternoon heat. And the verifiers that many Dark Factory teams built into their factories. When agents can challenge each other's work during the run, the system catches mistakes no single agent would.

HeatCrew's demo at the Universal Hackathon: told to ignore the heat, the safety agent refuses and hands the call to a human.

A human in the room

At the Universal Hackathon, nine of the 11 BAND projects include a human approval step: for a prescription, a ban, an order over budget, or a payment. At the Builder Base Buildathon, the BAND bonus made it a rule that no outreach goes out until a person approves it in the room. Teams didn't treat the human as a fallback for when the script breaks. They gave the human a seat where the work was already happening.

Rooms as the unit of work

Teams opened a room per incident (RecallRoom), per supplier negotiation (CapyBargain), per guest case (H.A.I.N), or per founder (First100), and treated its history as the audit trail.

Thanks

Thank you to WeAreDevelopers, lablab.ai, Docker and Featherless AI, HackerSquad and The AI Conference, ZooWork and AI Valley, Crewbase Collective, Tech Events SF, BridgeUs, Devnovate, Flexprice, and the AWS Builder Loft.

And most of all, thank you to every developer and team who built something with BAND over these two months. About one in three teams picked BAND.

Your coding agents already work on their own. BAND desktop lets them work as a team. Put them in one room and start with the AI Software Factory guide.