Play video
Ignacio Martinez registered the domain for this workshop on the Saturday before he gave it, and then asked the room not to request the biggest machines on offer, because he was paying for everyone's GitHub Codespace out of his own pocket.
Play video
Exa never meant to sell an API. Somebody sent them a message on Twitter asking for programmatic access to their search engine, and the answer was no, because there was no API to give.
Play video
An agent searching a contract for "30 days" has no way to know whether it has found a deadline, a grace period, or a retention rule, or wandered into a medication schedule entirely. Benjamin Clavié uses that to explain why coding agents turned out to be the easy case.
Play video
Somewhere between 30 and 50 percent of an agent's tokens get spent on searching, almost all of it up front, before any of the work you asked for. Maximilian David Rumpf treats that number as the whole opportunity.
Play video
Midam Kim spells her first name for a voice agent, letter by letter. The bot reads it back with an N on the end. She corrects it. The bot replies thank you for your correction, then says her name wrong anyway, because its pronunciation rules are English and her name is not.
Play video
A background agent quietly makes the tool call, then drops the result into the main model's context so the model believes it made the call itself. That sleight of hand is one of three tricks Bohan Li uses to run a voice agent on a slow, genuinely intelligent model without the caller noticing the wait.
Play video
The Moviefone hotline was a speech to speech system. The GPS unit in your parents' car was an event to speech system.
Play video
On a phone call with someone close to you, as much as 20 percent of the time you are both talking at once. Every real time voice model shipping today is half duplex: it is either listening or speaking, never both. That gap is the whole argument of this talk.
Play video
Sumanyu Sharma showed up for a doctor's appointment a voice agent had told him was booked. He was not on the schedule, the front desk turned him away, and he lost two hours.
Play video
When Andrew Qu handed Vercel's internal data science agent to its first trusted users, the verdict came back that it was awful. It had been clearing thirty percent of his evals and he thought the team was cooking.
Play video
Given the same fixed budget of roughly 600,000 tokens, an agent that did nothing but execute scored 76 on a bench of financial analysis tasks. An agent that spent part of that identical budget asking a second agent for advice scored 89. Same tokens, different jobs, better answer.
Play video
A research agent asks a human to approve its plan, then waits. The wait might last a month, spanning restarts and redeploys, and while it waits the function consumes no serverless execution time at all.
Play video
By the third version of the same agent, the source directory is gone. What remains is an AGENTS.md holding the system instruction and a small bash script that installs the GitHub CLI on first run. No tool definitions, no JSON schemas, no Python.
Play video
Mike Chambers puts a dictionary entry on screen: a harness is a set of straps and fastenings used to control an animal. Swap animal for model and the definition holds. His working version is subtraction. Take an agent, remove the model, and everything left over is the harness.
Play video
Half of Kay Malcolm's team sits in Europe and half in the United States, so when the Netherlands side commits code at four in the morning her time, the Americans wake up to the code and none of the reasoning behind it. Git records what changed, not why anyone decided it.
Play video
The Dioxus team got excited, maxed out their coding agent subscriptions, and turned out tens of thousands of lines of Rust covering features they had wanted for years. Almost none of it cleared the bar for merging. Those lines sat in draft, and Jonathan Kelley says they are sitting there still.
Play video
Because only the agent can reach an MCP server, a naive file upload tool makes the agent retype the file into the server character by character. Jeremiah Lowin calls it the world's most expensive copy paste operation, and it is the problem MCP apps exist to remove.
Play video
One of Max Drake's colleagues turned a whiteboard app into a window manager, then used it to play Pong with real desktop windows. Drake is a product engineer at tldraw, the London company behind the free infinite canvas app and, more importantly to him, the SDK powering other people's canvases, including Replit's.
Play video
In one of the demos an agent hits an error, decides the fix is to delete the table and start fresh, and does exactly that. Nothing stops it.
Play video
Vincent Wendy pointed an agent at the sponsor banner in the lobby, a wall carrying more than 140 logos, and asked it which ones were missing. It caught every omission in the checks he ran, and he used the same trick on the conference shirt.
Play video
Alex Hancock built one of the clients in this talk the night before he gave it, which is roughly the point. Hancock is a software engineer at Block, works on the open source agent harness Goose, now donated to the Linux Foundation, and maintains the Rust SDK for MCP.
Play video
Adding a nice interface made the product worse. With plain text results the model would run ten or fifteen job searches, filter them, and assemble a table. Once a rendering widget was attached, it called the tool once, saw results already on screen, and stopped exploring.
Play video
A single token of KV cache on Mistral 7B costs 131 KB. Multiply that by 16,000 tokens of context and 80 concurrent users and the cache alone wants 42 GB of GPU memory, which is why requests start failing on a 24 GB card.
Play video
Ben Guo built the page listing every speaker in this session while he sat waiting to go on, then put a QR code to it on his first slide. He builds that way because his computer is a server in the cloud that he talks to in plain language.