August 26-27, 2026 · Cloudflare SF, 101 Townsend · 40 builder seats
Applications opening soon · 40 of 40 builder seats left
The Agent Natives
Builders Hackathon
Two days where forty builders from startups that already ship make their products legible to AI agents. Not chatbots bolted onto human websites, but products designed for a web where agents are first-class users.
The event itself runs on that stack. Your application, your team, your submission and your score are all agent tool calls. The page you are reading is a client of the same API your agent will use.
Startups with something already shipping.
This is an invited cohort, not an open call. We are looking for founders and engineers who have a product in market and two days to spend making it agent-native: an MCP server, agent-readable surfaces, machine-to-machine auth, agent payments.
The reason for the filter is simple. The infrastructure decisions you make that week are the ones your company keeps. We would rather have forty of those than three hundred weekend prototypes.
Forty builder seats. The room at 101 Townsend seats fifty with tables, and ten of those chairs are held for judges, mentors, partner engineers and the Cloudflare host who stays on site both days. Forty is what is left, and forty is the number that is real.
You apply, a human reads it, and the seat is yours only when approved. Applying does not hold a seat, which is why a full queue never locks out the room. Applications close Friday, August 21, because the roster has to freeze before Cloudflare prints badges and sends the NDA.
Two doors. The agent one is faster.
If you already have an Immersive Commons agent token, your agent can fetch the questions, write the application, record the venue NDA once Cloudflare sends it to you, form your team and pull the sponsor context pack without you touching a form. If you do not have one yet, minting it is one browser approval.
npx -y @immersivecommons/cli auth login --scopes hack:read,hack:register,hack:team,hack:submitThen hand your agent the MCP endpoint https://www.immersivecommons.com/api/mcp and point it at anb-hack-01. It calls ic_hack_application_form to get the questions for your applicant type, then ic_hack_apply with your answers. Every question carries a why field saying what the reviewer is looking for, so a good agent answers the question behind the question.
One thing to know before it writes: if you give us a product URL, we fetch it and record which agent surfaces it actually serves. Finding none is not a mark against you.For most of this cohort “none yet” is the honest answer and the reason to come. A claim the fetch contradicts is worse than an honest gap.
Then a person reads it. There is no automatic approval and no queue position to game: your seat exists the moment a reviewer says yes, and not before. Applications close Friday, August 21.
Prefer not to use an agent?
There is no web form, and we would rather say so than point you at a second door that half works. Write to admin@immersivecommons.com with what you are building, its URL, and what you would spend the two days on, and we will file the application on your behalf. It lands in the same queue, in front of the same reviewer, under the same deadline. Nobody is disadvantaged for arriving the human way; you will simply do by hand on the day what the agent path did in advance.
The floor's whole toolkit, for two days.
- Model access. A key for the weekend, requested through the same token your agent already holds.
- Agent rooms. A live multi-agent collaboration room per team, on the Cotal mesh.
- The corpus. Query the floor's research index of papers, talks and prior builds from inside your build.
- A context pack. One folder link admits your agent to the whole subtree: sponsor docs, stack guides, the judging rubric.
- Judges who ship. Panel drawn from the venue, the sponsors and the floor.
- Somewhere for it to live. Your project keeps a profile after the weekend rather than dying as a repo nobody revisits.
Agents that build with you, not just for you.
Builder agents run on your machine, with your keys. You bring your own agent and it joins the room's mesh through a scoped, expiring grant. It has a shell because it is yours and it is on your laptop. We supply the room, the grant, the brief and the context. Not the credentials.
What your agent reaches through that grant is the floor's read surface: the research corpus, the sponsor context pack, the rubric, and the event's own tools. Enough to argue with your architecture, nowhere near enough to run it.
Two things we are not offering, said plainly rather than left to be discovered on the day. We are not hosting mentor agents you can summon into your room; that lane got cut so the front door could work, and we would rather cut it than ship it thin. And we are not handing forty strangers a sandboxed-enough shell on our infrastructure, which is the kind of thing that sounds fine until it isn't. Both are in the lab. Neither is in this event.
Two full days, and a hard stop.
| Wed 9:00 | Doors, check in, badge pickup |
| Wed 9:30 | Kickoff: rules, judging rubric, sponsor tracks |
| Wed 10:00 | Cloudflare stack session |
| Wed 11:00 | Team formation, build begins |
| Wed 13:00 | Lunch, catered |
| Wed 18:30 | Dinner, catered |
| Wed 20:30 | Building empty |
| Thu 9:00 | Doors, build resumes |
| Thu 15:00 | Submissions lock |
| Thu 15:30 | Demos to the judging panel |
| Thu 18:00 | Results, awards, closing |
| Thu 20:30 | Building empty |
All times PDT. Doors are 9:00am both days and the venue clears at 8:30pm both nights. The clear-out is not a soft close, so we schedule to finish an hour before it. This is a two-day sprint, not an overnight one, so bring a sleep schedule.
Submitting is a tool call. So is scoring.
Your team has one submission and you overwrite it as often as you like until it locks at 3:00pm on Thursday, which is what “I fixed the demo link at 2:55” actually means. The field that carries the most weight is not what your product does, it is what makes it agent-native.
Standings rank by the mean across judges rather than the sum, so a team seen by two judges is not quietly beaten by a weaker team seen by four. Judge notes come back to you after results.
The rubric
One hundred points, and the same shape in every track: 25 that it runs, 30 for the track’s agent-native criterion, 20 for the depth criterion that track cares about, 15 for landing in the real product, 10 for the demo.
“It runs” is a gate as well as a criterion. A submission a judge cannot trigger live on the day cannot place, whatever it scores everywhere else. A broken agent-native concept is a slide deck, and a room of builders who each gave up two working days deserves a winner whose thing actually ran.
Above that floor the weighting is deliberate. Once the software runs, the only question worth thirty points is how much of the product’s value an agent can reach with no human in the loop. And the fifteen for landing in the real product is there because everyone here arrives with something that already has users, which makes the usual hackathon ending, a beautiful weekend fork nobody merges, both avoidable and unforgivable at this event.
Track 1: The Front Door
Your product should be completely usable by an agent that has never seen it, found it on its own, and has no human sitting next to it.
- Cold-start success, 30. An agent nobody on the team briefed completes the task from discovery alone. Judge-triggered, live, on the day.
- It runs, 25. Real product, real data, real credential. Not a mock, not a recording, not a localhost demo that cannot survive the venue wifi.
- Surface quality, 20. Scoped and revocable auth, errors that tell an agent what to do next, write gates on consequential verbs, discovery docs that match the actual tool list.
- Lands in the product, 15. Merged, deployed, and reachable after the weekend, with a rollback path. A branch nobody will merge scores zero here.
- Demo, 10. Working software inside thirty seconds, honest about what does not work yet.
Track 2: The Colleague
Your product stops being a place people visit and becomes a party that other people's agents delegate to, negotiate with, and remember.
- Real work across a real boundary, 30. The counterparty is not yours, the exchange changes state on both sides, and the contract was not privately agreed by two people at the same table.
- It runs, 25. Live, on the day, with the other side present. Not a transcript.
- Coordination design, 20. What happens when the other agent is offline, slow, wrong, or new. Durable delivery, presence, retries, and memory that survives the session.
- Lands in the product, 15. Merged, deployed, reachable after the weekend, with a rollback path.
- Demo, 10. Working software inside thirty seconds, and the failure case shown on purpose.
Track 3: The Receipt
An agent-native product does consequential things on someone's behalf, so the build is the proof: what it may do, what it refused to do, and what actually happened.
- Provable safety and correctness, 30. A gate that refuses in the tool layer, a suite that has actually gone red on a real regression, and a receipt an outsider can verify.
- It runs, 25. Live, on the day, against your real product. The refusal happens on stage, not in a screenshot.
- Adversarial survivability, 20. A judge attacks it live. Injection, scope escalation, replay, budget overrun. Surviving one you did not anticipate scores highest.
- Lands in the product, 15. Merged, deployed, enforced in production, with a rollback path.
- Demo, 10. Working software inside thirty seconds, and an honest account of what it still cannot stop.
Agent pre-reads may help the panel triage. They do not decide. A human picks the winner and puts their name on it.
Three organizers, one venue host.
Immersive Commons. 300+ member builder community at Frontier Tower SF. Runs Vibe Coding Nights and the Web for Agents series, and operates its own agent-native membership platform, the one this event runs on.
Runtype. Platform for building, running and shipping production AI agents, flows and agent-powered products.
Cotal.ai. Open standard protocol for AI agents to collaborate in shared spaces, from peer-to-peer swarms to hierarchical teams.
Cloudflare hosts us at 101 Townsend and runs a session on their stack. Supporting partners: Nebius, Z.AI, hirey.ai, Mitosis Labs and thesingularity.so. Community: Learning Layer Labs, ship.builders, Dabl Club, ClawCamp and Weba.network.
The logistics nobody likes but everybody needs
- Every attendee signs the venue NDA. Cloudflare sends it to the confirmed roster 48 hours ahead. Unsigned means you cannot be checked in, and that is enforced by the check-in tool rather than by goodwill at the door.
- Applications close Friday, August 21. Badges are printed in advance, so the roster freezes days before the doors open and late approvals are not possible.
- Meals both days are covered.
- Doors 9:00am PDT. Building empty by 8:30pm PDT. Both days.
Forty builder seats.
Bring a product that already exists. Leave with one an agent can use. Applications close Friday, August 21, and a person reads every one.
Sponsor, judge or volunteer enquiries: get in touch. Those roles are granted by an organizer and come with their own tools.