AI Engineer World’s Fair / San Francisco / 2026

A field guide
to what changed.

The public recordings, mapped into nine currents and distilled for curious builders who would rather follow an idea than a playlist.

Explore the currents Give me a watch path

Topicmap

Updated August 14, 2026June 29 – July 2, 2026Official event ↗

FIELD NOTE / 01

The model is no longer the whole product. Useful agents are built in the layers around it—the context, tools, permissions, and evals that shape what it can do, plus the human judgment that decides what is worth doing.

Curated / Start here

Four talks to understand the moment.

A recommended sequence for an engineer who wants the ideas beneath the announcements: capability, practice, understanding, then ambition.

Gaurang’s starting shelfThese are the talks I’d hand to a curious builder first. Together they explain what changed, how teams are adapting, and what humans still need to get better at.

01 · Capability01 / 04

Field Guide to Fable

Thariq Shihipar, Claude Code, Anthropic

Start here for a practical reset on how to work with models whose abilities have outgrown the old prompting playbook.

What you’ll take away

Give the model tools and let it assemble context instead of pasting everything into the prompt.
Stronger models often need smaller prompts, fewer examples, and fewer prohibitions.
Make unknown decisions visible: use orientation passes, interviews, prototypes, and an explicit unknowns log.

02 · Practice02 / 04

Simon Willison in conversation with Cat Wu & Thariq Shihipar, Anthropic

Cat Wu & Thariq Shihipar, Anthropic · moderated by Simon Willison

The broadest snapshot of how an AI-native engineering team is actually changing product work, review, memory, and safety today.

What you’ll take away

The developer’s centre of gravity is moving from manual implementation toward product sense and steering.
Proactive, shared agents introduce a multiplayer workflow: team memory, background work, and agent-authored pull requests.
More autonomy needs stronger evals, red-teaming, permissions, and trusted execution—not less engineering discipline.

03 · Understanding03 / 04

Understanding is the new bottleneck

Geoffrey Litt, Notion

A vital counterweight to the autonomy story: if agents write the code, engineers need better ways to build understanding—not permission to stop learning.

What you’ll take away

Understanding is not only for verification; it is how humans remain creative participants in the next decision.
Turn every substantial diff into a tailored explainer with intuition first, interactive figures, and literate code.
Use quizzes and temporary ‘micro-worlds’ as speed regulators so delivery never outruns comprehension.

04 · Ambition04 / 04

Everything we knew about software has changed

Theo Browne, @t3dotgg

Finish here for the strategic provocation: when implementation collapses in cost, the projects that once felt unreasonable become the new starting point.

What you’ll take away

Do not preserve developer rituals merely because they are familiar; question the skeuomorphic software workflow.
Work has shifted tiers: some startups become side projects, and some side projects become a Markdown file on a schedule.
Use the new leverage to build across a wider product surface, not only to make the old scope incrementally faster.

01 / Topic map

Follow the currents.

Nine recurring bodies of work emerged across the fair. Pick a current to filter the full archive.

02 / Watch paths

Four routes through the fair.

Don’t know where to begin? Choose the question you’re trying to answer and follow a four-talk path.

01 · Start here

The new shape of software

Four talks for understanding why code generation is no longer the interesting part.

02 · Build

Ship an agent that survives

From the open harness to evals, verification, and production-grade deployment.

03 · Go deep

Context is the product

The emerging stack for memory, retrieval, compaction, and grounding.

04 · Lead

Build the AI-native team

A practical route through product ambition, embedded engineering, and attention.

03 / Full archive

Find your next idea.

⌘ K

Sort

Showing 18 talksAcross all currents

Agent runtimesDeep note

Frank Coyle, UC Berkeley

Frank Coyle places typed entities, relationships, and ontology constraints outside the probabilistic model as logical guardrails. Validating proposed tool calls with Pydantic and their effects against RDFS or OWL catches duplicate refunds, invalid states, and confused roles that prompts alone miss.

Product & designDeep note

Thariq Shihipar, Claude Code, Anthropic

Anthropic's Thariq Shihipar launches Fable with a field guide: unhobble the model with tools, hunt your unknowns, grieve the old craft, and get unreasonable about ambition.

Product & designDeep note

Theo Browne, @t3dotgg

Theo Browne (t3dotgg) argues that as models jump from tool-calling to orchestration, the answer is not to get better but to go bigger. He walks through his own "AI psychosis," the skeuomorphic habits engineers cling to, and the way every tier of software has shifted down — startups becoming side projects, and side projects becoming a single markdown file. His closing challenge: if your idea doesn't feel stupid, it isn't big enough.

Coding agentsDeep note

Geoffrey Litt, Notion

Geoffrey Litt argues that understanding is not merely a correctness check—it is how engineers remain creative participants. His practical toolkit turns agent-written changes into explainers, quizzes, interactive figures, and temporary micro-worlds that keep delivery from outrunning comprehension.

Agent runtimesDeep note

Frank Coyle, UC Berkeley

The Claude Certified Architect exam hands you six production scenarios and picks four at random, and Frank Coyle walks through them backwards, leading with the anti pattern in each one. Knowing what not to do is what points you toward what to do, the same way the design patterns movement of the early 1990s came with a catalog of the moves that quietly ruin you.

Product & designDeep note

Garry Tan, Y Combinator

Y Combinator president and CEO Garry Tan makes the case for the AI-native company. He argues leverage comes not from the model but from how you wire the work: skill files as employees, a resolver table as an org chart, and a curated company brain that compounds. He backs it with YC batch data and his own open-source second brain, GBrain.

Agent runtimesDeep note

Lance Martin, Anthropic

Lance Martin's long-horizon harness separates a stateless “brain” from sandboxed “hands” backed by an append-only session log, pairs builders with independent verifier contexts, and consolidates useful execution traces into reusable skills so asynchronous agents can recover, self-check, and improve securely.

Agent runtimes

Practical guidance for designing reusable skills that make agents more capable and dependable.

Models & inferenceDeep note

Alexander Embiricos & Romain Huet & Peter Steinberger, OpenAI

Alexander Embiricos and Roman Huitt of OpenAI argue that far from replacing engineers, AI is triggering a return to engineering's roots — "AI engineers are eating the world." They frame Codex as a layered, open stack (model, responses API, open-source harness, apps server, app-layer plugins) that OpenAI uses internally exactly as it ships to developers, and preview the GPT 5.6 series across cost-efficient (Terra, Luna), fast (Codex Spark, Cerebras at 750 tok/s), and frontier variants. The talk closes on "value maxing" — extracting real value from agents through cost, speed, and parallel cloud execution rather than "token maxing."

Evals & verificationDeep note

Annabell Schäfer, Langfuse

Annabell Schäfer shows why self-improvement loops need a precise target function, task-specific evaluation, and domain experts who know when automation has hit a wall. The practical design includes explicit handoffs and escape hatches so an agent collaborates with humans instead of burning tokens indefinitely.

Evals & verificationDeep note

Philipp Schmid, Staff Engineer, Google DeepMind

SkillsBench found almost no evals across 50,000 GitHub skills. DeepMind-style CI on skills, plus lean descriptions and negative cases, is how preference skills stay honest.

Context & memoryDeep note

Tim Ainge, Good Collective

Tim Ainge gives a practical tour from extracting graphs out of unstructured text to schema and ontology design, personalized PageRank, shortest paths, and subgraph matching. The throughline is using graph-native and hybrid algorithms to make AI applications smarter, cheaper, and more reliable.

Agent runtimes

Abhishek Bhardwaj, OpenAI

Abhishek Bhardwaj describes designing a cloud sandbox system for safely running many agent processes at scale.

Agent runtimesDeep note

Gagan Bhat & Isabella Kai He, Anthropic

Gagan Bhat and Isabella Kai He show how Anthropic decoupled an agent’s reasoning loop from its tool-execution environment, made sessions durable and recoverable, and used logs for context recovery and ‘dreaming’ as model capabilities outgrew earlier harness assumptions.

Product & design

Addy Osmani

Addy Osmani argues that choosing what is worth building matters more than raw execution speed.

Models & inferenceDeep note

Nader Khalil, Director of Developer Technology, NVIDIA · Joseph Nelson, Cofounder, CEO, Roboflow · Alex Cheema, CEO, EXO Labs · Ahmad Osman, Founder & CEO, Osmantic · Matthew Berman, Founder, Forward Future

Five local-AI insiders from NVIDIA, Roboflow, EXO Labs, Osmantic, and Forward Future argue on-device models just crossed from toy to default for privacy, cost, and control.

Context & memoryDeep note

Stephen Chin, Neo4j

Stephen Chin compares vector and graph memory on the same home-network data: only the graph agent follows multi-hop relationships to find obsolete software and exposed ports. Vector-seeded traversal makes large-scale memory precise, explainable, and auditable without loading every markdown file.

Coding agents

Kyle Mistele, HumanLayer

Kyle Mistele revisits coding-agent loops from first principles, questioning lights-off software factories where people interact only through tickets and specifications and no longer read the code.

Method / Read this

A map, not an official archive.

This independent guide indexes public conference-talk uploads from the official AI Engineer YouTube channel through August 14, 2026. It is a curated snapshot, not a claim of an official or final talk count.

153 “Deep note” summaries are recording-informed from official YouTube transcripts. The remaining concise notes are catalog-informed interpretations of public titles and metadata—not claims of full transcript review.

Official YouTube channel Official event page Video index & method Transcript field corpus

A field guide
to what changed.