November 11–12, 2026 · San Francisco
The people building coding agents, software factories, and the tools around them. Explore their illustrated reading editions, recordings, and transcripts.
50 of 50 sessions · Event order · timestamps are positions in each source recording
Gimlet Labs
November 11 · Main stage
Natalie SerrinoGimlet LabsCo-Founder
Natalie Serrino asks her agent to make a slide look better. It does. The result takes less time than she would have needed to redesign it herself, and she gets a slide she likes without becoming a designer.
Anthropic
November 11 · Main stage
Lydia HallieAnthropicMember of Technical Staff
A setup built around an older model can become an obstacle for a newer one.
Cloudflare
November 11 · Main stage
Michelle ChenCloudflareGroup Product Manager
Cloudflare built a distributed inference network around a bet: models would get smaller and faster, making the distance between the model and its user increasingly important.
SpaceXSI
November 11 · Main stage
Lauren TanSpaceXSIPrincipal Engineer
When an agent makes the same mistake again, it is tempting to add another warning to AGENTS.md. Lauren Tan wants the correction to land somewhere harder to ignore: the codebase, an API, or an executable check.
Intrinsec AI
November 11 · Online track
Speaker identity unverifiedIntrinsec AI
A guardrail can inspect a prompt before the model receives it and inspect a response after the model produces it. But between those checks, the model has already transformed its input into internal state. The talk asks whether an inference operator can use that state as an additional security signal.
Harvey
November 11 · Online track
Joey WangHarvey
Joey Wang opens with the scale of Harvey's internal agent platform: more than one agent starting every minute on average, around the clock, and thousands of agent-written pull requests.
Amazon Web Services
November 11 · Online track
Rajath KotyalAmazon Web Services
An e-commerce specification says that an order should become a backorder when inventory is unavailable. Elsewhere, it says that no order should ever become a backorder.
Meta
November 11 · Online track
Chinmay GandhiMeta
Swaapnika MotapothulaMeta
Vlad TsvangMeta
Twelve alerts arrive at two in the morning. An assistant can help explain a latency spike, inspect a deployment, and read a runbook. Someone still has to decide which alert matters, ask each question, connect the answers, and carry out the follow-up.
Google DeepMind
November 11 · Main stage · Excerpt
Philipp SchmidGoogle DeepMindMember of the Technical Staff
An agent can reach the right answer while taking an unnecessarily expensive route. Philipp Schmid's harness experiments focus on that difference: two systems may look equivalent in an accuracy table, yet consume very different numbers of turns, tokens, and dollars to finish the same work.
Databricks
November 11 · Main stage
Ankit MathurDatabricksMember of Technical Staff
A broken data pipeline invites several plausible explanations. Perhaps an upstream job changed. Perhaps the data no longer matches an assumption. An agent can suggest each theory, but testing it against production may be the very thing an engineer cannot safely let it do.
Sonar
November 11 · Main stage
Andrew FinchSonarStaff Software Engineer
A coding agent makes it easier to produce a pull request. That does not settle whether the change should merge, whether it behaves correctly after deployment, or whether a vulnerability discovered next week changes the answer.
Factory
November 11 · Main stage
Luke AlvoeiroFactoryMember of Technical Staff
A team hands an agent a large goal, approves the plan, and returns later to completed work.
Cline
November 11 · Online track
Ara KhanCline
A coding agent can help build the system that makes coding agents better. That observation is easy to turn into a story about a distant intelligence explosion. Ara Khan instead begins with a nearer question: what should an engineer work on when a new model can erase the value of a side project overnight?
Tanmay Sah
November 11 · Online track
Tanmay Sah
Dolly Sah
The coding agent has done what it was asked to do. It patched the repository, passed the tests, and produced a release. Then a downstream integration fails. Reverting the Git commit sounds like an obvious response, but the repository is only one part of the system the agent changed.
Duolingo
November 11 · Online track
Kush AgrawalDuolingo
Duolingo's iOS codebase contains roughly 1.8 million lines of code and a decade of accumulated decisions.
Zenity
November 11 · Online track
Roey Ben ChaimZenity
An agent reads cloud configuration, queries identity services, and probes instance metadata.
TypeSafe AI
November 11 · Main stage
Diogo AlmeidaTypeSafe AICo-founder & CEO
A model can be impressive in a demo and still be unsuitable for unattended operation. Diogo Almeida makes that gap the center of his keynote. He likes the new tools; his objection is to treating a good assistant as if it were already a dependable automation component.
Glean
November 11 · Main stage
Chinmay GoyalGleanSoftware Engineer
A search assistant makes many decisions that do not require it to write anything.
UiPath
November 11 · Main stage
Chibi VikramUiPathEngineer
Kevin BernsteinUiPathPrincipal AI Solution Architect
A workflow does not need to deliberate about every step.
ZoomInfo
November 11 · Main stage
Brian KennedyZoomInfoPrincipal Engineer
An agent says it scheduled an email for Tuesday.
Workday
November 11 · Main stage
Scott RobertsWorkday
Workday users do not always arrive with an open-ended research problem. Sometimes they want their pay information, a time-off action, or a colleague's contact card. Scott Roberts asks what happens when an interface can recognize that intent quickly enough to prepare the action while the user is still typing.
Microsoft
November 11 · Online track
Girish KondaMicrosoft
An orchestrator delegates work to several agents.
Nikolay Petrov
November 11 · Online track
Nikolay Petrov
A skill can teach an agent something it does not know, or it can steer the agent through a process it already understands.
Lovable
November 11 · Main stage
Fabian HedinLovableCTO
Lovable nearly failed by building for agents too early. Fabian Hedin opens with that admission, not with the architecture diagram. In 2023, the team could see a future of agents delegating to agents, but the models available at the time could not reliably support the elaborate system they built around them.
Ramp
November 11 · Main stage
Zach BruggemanRampPrincipal Software Engineer
Inspect began as Ramp's collaborative coding agent.
Independent
November 11 · Main stage
Steve YeggeIndependentIndependent Advisor & Creator of Beads & Gas Town
Steve Yegge has spent the year trying to make agents run other agents.
Independent
November 11 · Main stage
Simon WillisonIndependentCreator of Datasette
Simon Willison organizes his year-in-review around three inflection points: coding agents becoming everyday tools, models becoming capable of pursuing much larger goals, and generated games becoming genuinely enjoyable to him.
OpenAI
November 12 · Main stage
Jason LiuOpenAIDeveloper Experience
A thousand agents running in the background is an impressive quantity. It does not tell Jason Liu whether their owner is doing ambitious work. He wants to know what the person is trying to make, whether they care about the result, and whether they can recognize when it needs another pass.
Earendil
November 12 · Main stage
Armin RonacherEarendilCTO
Ronacher's coding agent has become much more capable while its core harness has barely changed.
Warp
November 12 · Main stage
Zach LloydWarpFounder & CEO
Running several coding agents on a laptop changes how much work one developer can attempt.
ArtCraft
November 12 · Main stage
Brandon ThomasArtCraftCEO
Thomas arrives after a month in which a project he had not yet posted to Hacker News escaped into the wider internet.
Ayush Bhardwaj
November 12 · Online track
Ayush Bhardwaj
One agent downloads a documentation page and keeps forty lines. Another parses a dependency report and keeps the tail because that is where the totals happen to live. A third runs a test suite and truncates the output. Each gets an answer by writing a small program around the program it was asked to use.
Brian Douglas
November 12 · Online track
Brian Douglas
Douglas can ask an agent to write Rust, but he does not claim the same Rust judgment as his more experienced teammates. The code may compile and receive a positive automated review, then attract substantial pushback from a human who understands how the company wants its systems to fit together.
Rohan Pandey
November 12 · Online track
Rohan Pandey
A model's downloaded weights do not tell an operator which serving configuration will work best for a particular machine and workload.
Maersk
November 12 · Online track
Dmitry BuykinMaersk
Buykin once dealt with an expanding AI-written project by deleting unused tests and boilerplate, then rewriting the whole thing. Useful pieces disappeared with the mess. Worse, the rewrite did not solve the process that created it: agents kept adding code, and complexity grew back.
Cognition
November 12 · Main stage
Walden YanCognitionChief Product Officer
Cognition's merged pull requests grew much faster than its headcount. Yan opens with that internal comparison because it creates a more interesting question than whether an agent can write code: what does an engineering organization have to change to absorb all that work?
Exa
November 12 · Main stage
Jeffrey PinnerExaCTO
A model can reason well and still lack the fact that makes its answer correct. Pinner opens with the growth of search traffic at Exa, then turns to the more durable reason agents need search: the world changes after training, and training never contained all of the world in the first place.
AWS
November 12 · Main stage
Doug ClausonAWSPrincipal Product Manager, Kiro
Few people raise their hands when Clauson asks whether their teams are getting everything they expected from AI. His diagnosis is that many teams have attached an agent to an existing process and accelerated only the code-generation step. The rest of software development still runs at its previous speed.
Anthropic
November 12 · Main stage
Lance MartinAnthropicMember of Technical Staff
Formalizing a mathematical proof is different from writing a convincing explanation.
Syntax
November 12 · Online track
Scott TolinskiSyntaxCo-host
CJ ReynoldsSyntax
Scott Tolinski and CJ Reynolds spend their lunch session moving from state machines to programming languages, native applications, skill loss, and a game about identifying model-generated interfaces.
Robinhood
November 12 · Online track
Chetas JoshiRobinhood
A pull request widens a column from integer to big integer. CI passes. The registered downstream consumers look fine. Is it safe to merge?
Anthropic
November 12 · Main stage
Jarred SumnerAnthropicFounder of Bun
Jarred Sumner wanted Claude Code to use less memory and run faster. One possible answer was to rewrite it in Rust. He had experience with that kind of rewrite, but the tests changed the decision.
Uber
November 12 · Online track
Preetam DwivediUber
Manjari AkellaUber
Two pull requests pass CI against the same base. Merge the first, then the second, and the combined branch breaks. Neither green result tested the state that ultimately reached the trunk.
Google DeepMind
November 12 · Main stage
Ishaan DeyGoogle DeepMindSoftware Engineer
When building a prototype becomes cheap, it changes when a team can learn. Instead of spending most of its effort imagining the right interface before implementation, the team can make several versions, try them, and discover details it could not have reasoned out in advance.
Diffusion
November 12 · Main stage
Justin McCarthyDiffusionCEO
Justin McCarthy opens a software-factory talk by asking people to stop building software factories. The provocation has a scope: stop forcing uncertain ideas through an industrial process before you know what you want.
Independent · Neo4j
November 12 · Online track
Stephen ChinNeo4jVP of Developer Relations
Cassandra ChinIndependentBook Author, Keynote Speaker, and Kids Workshop Instructor
A pull request can contain good code and still give its reviewer a bad account of what happened.
Google DeepMind
November 12 · Main stage
Melvin JohnsonGoogle DeepMindVice President, Gemini Post-Training Lead
An engineering task stops being useful training material when every attempt succeeds.
Amp
November 12 · Main stage
Quinn SlackAmpCEO
Quinn Slack's obstacle was not getting permission to use an agent. He was the CEO. The difficulty was changing a company whose existing customers, roles, and operating habits had been built around something else.
Factory
November 12 · Main stage
Dexter HorthyFactoryMember of Technical Staff
A software factory already existed before coding agents. Someone identified a problem, translated it into work, built a change, ran tests, reviewed it, released it, and watched what happened. Dexter Horthy's argument is that replacing the builder does not make the rest of that system optional.
Independent
November 12 · Main stage
Thariq ShihiparIndependent
Thariq Shihipar starts with predictions he got wrong. Agents did not settle into one permanently correct tool interface. Work he expected to stay inside familiar web frameworks expanded into other languages and applications. Capabilities kept moving beyond the boundary he had drawn around them.
