Talks

Agents have already taken over writing code, and I think review, maintenance and deployment are next. What's left is deciding what to build and defining what good looks like, which is the job of a product engineer, and I think that's what most of us are about to become.

We built memory for Alyx, Arize's in-product agent, the way the industry recommends, with a knowledge graph and hybrid search. Then we shipped a single structured file per user instead. This is why, and how we test that it actually helps.

Agents made writing code cheap, so the bottleneck moved to reviewing it, and reviewing harder doesn't work for humans or for LLMs. I think the human job moves up a level, to writing the specs, tests and evals that the code gets checked against.

In 2025 models started dropping instructions at around 200. I re-ran the benchmark on 2026 models and they handled thousands, so the problem is no longer how long your prompt is but whether the model actually follows it.

A hands-on workshop where we trace a deliberately bad financial analysis agent, read the traces, write code and LLM-as-judge evals, check the judge against human labels, and run experiments to fix it. I've given it in Python and TypeScript, with Phoenix and with Arize AX.

The conference-length version of the workshop. If you're checking your agent by trying a few prompts and eyeballing the output, you can't tell whether a change made it better or worse, and evals fix that. The Vector Space Day version applies the same idea to retrieval.

A head-to-head experiment: the same 25 GitHub tasks done by an agent using the GitHub MCP server and by the same agent using command-line skills, comparing cost, speed and correctness, and when each one is the right choice.

How we improved Claude Code's score on SWE-Bench Lite by up to 11% just by changing its CLAUDE.md, using written feedback from evals instead of a numeric score.

What Arize learned building and evaluating Alyx, our in-product agent, in production. Most of the lessons are about enforcing behavior in code instead of asking nicely in the prompt.

An introduction to the Model Context Protocol for San Francisco city government: what it is, how it works, and the ethical, environmental and economic problems that come with AI.

Five LlamaIndex customer stories about agents in production, most of which turned out to be about document parsing.

A lightning fast overview of the agent to agent protocol communication space, which in May 2025 had at least 12 competing protocols. Probably my most-watched talk ever, for some reason.

Is AI going to replace programmers? I go through the Industrial Revolution, the history of abstraction in software and the economics of programming to argue that it's going to create a lot more programmers, while getting rid of a lot of jobs along the way.

My most-repeated talk of 2025: how to build agents with LlamaIndex Workflows, organized around the common agent design patterns (chaining, routing, parallelization, orchestrator-workers and evaluator-optimizer). I gave versions of it everywhere from hackathons to Stanford.

Short introductions to LlamaIndex for hackathons and meetups, from 30 minutes down to 5, with advice on what to build.

How to run multi-agent systems in production with llama-agents and later llama-deploy, treating each agent as a microservice.

Why naive RAG falls over on summarization, comparison and multi-part questions, and how agentic techniques like routing, memory, planning and tool use fix it.

How retrieval-augmented generation works (ingestion, indexing, storage and querying) and 7 advanced querying strategies to make it better. I gave it a lot, mostly as webinars and meetups with database and infrastructure companies.

A model of the 9 stages every web technology goes through, from wild experimentation to mass adoption, which I used to predict what comes next for frameworks, serverless and, by 2023, AI.

The results of the Jamstack Community Survey, which I ran at Netlify every year from 2020 to 2022.

What 7,500 web developers told the Jamstack Community Survey about how they actually build websites: frameworks, rendering, serverless, and the gap between the best practices people say they follow and what they actually do.

What npm download numbers say about framework popularity once you measure them relative to the growth of the whole registry, and why Vue is probably more popular than it looks. It started as a Twitter thread.

What the Jamstack actually is (a request pattern, an architecture and a workflow) and why serverless looks a lot like the move to virtual machines all over again.

The whole web from zero for brand-new bootcamp students: the Internet, DNS, HTTP, HTML, CSS, JavaScript, security and APIs, ending with deploying a site.

My opening keynote at JSConf EU, using npm's survey of 33,000 developers and the registry's download data to describe who JavaScript developers are, what they use, and where the language is going.

How security works in the npm registry, what real incidents like event-stream taught us, and what you can actually do to protect your apps.

Postmortems of the worst outages in npm's history, including left-pad, which I will feel genuinely bad about forever.

What enterprise and finance teams need to know about building with open source JavaScript safely.

What npm knows about how people use JavaScript, from the registry's download data and our developer survey, plus a series of ill-advised predictions about where it's going.

A history of JavaScript modules, and why modular JavaScript is how front-end development gets back under control.

What changed in npm 5: lockfiles, a client 10x faster, offline installs, npx, and a faster registry.

A comedy talk about all the ways engineers solve scaling problems they don't have, in 22 chapters in 22 minutes. The Thunderplains version has the swearing censored.

A roadshow for front-end teams at companies, about automating everything with npm, using small modules, and the best practices that go with them.

A tour of the npm features most people don't use: run scripts, scopes, orgs and teams, dist-tags and webhooks.

What JavaScript development looked like before npm, how npm got started, and where we wanted to take it, including becoming the package manager for front-end code.

How npm Inc used service-oriented architecture to scale both the registry and the team, and why a monolith is still the right choice for version 1.

A marathon, 2-hour overview of professional web development that I gave to graduates of Hack Reactor and other bootcamps more than 30 times between 2014 and 2021. It's now also a website.