AI Engineer World's Fair 2026
2026-06-29..2026-07-02
San Francisco, United States
[AI Engineer World's Fair 2024, AI Engineer World's Fair 2025, AI Engineer World's Fair 2026]
as the organisers put it
The largest technical AI conference in the world, with 29 tracks, 300 speakers, 100 expo partners, 6,000+ AI Engineers, founders, and VPs of AI. June 29 - July 2, 2026 • San Francisco, CA
| title | author | length |
|---|---|---|
| Recursive Coding Agents | Raymond Weitekamp | 24m |
Claims specific wins: RLMs let agents operate on external prompts as executable state and achieve real benchmark improvements — Qwen 3.59B as an RLM beats… | ||
| The Log Is The Agent | Ishaan Sehgal | 15m |
Argues that the agent is the append-only session log not the model or runtime and demonstrates how reconstructing state from that log enables durability… | ||
| Production Evals For Agentic AI Systems | Nishant Gupta | 08m |
Benchmarks score models 90% then 92% but production shows failures when plans call tools or long workflows run for hours. The claim is that continuous… | ||
| The Miranda Hypothesis: How Hamilton Poisoned Persona Evals | Jacob E. Thomas | 58m |
Claims that RLHF amplifies a corpus-driven failure mode called Miranda distortion and that standard persona evals miss it. Shows a reproducible instrument, the… | ||
| A Genius With Amnesia | Victor Savkin | 20m |
Claims that agents are constrained by repo boundaries and episodic amnesia and that this causes seven explanations for a single cross-repo change. Describes a… | ||
| Turn 10,994 Notes Into Memory | Paul Iusztin, Louis-François Bouchard | 40m |
A concrete system turns a 10,000+ note Obsidian vault into live agent context: three iterations over 18 months produce a three‑layer pipeline—raw files, an… | ||
| Build Systems, Not Code | Angie Jones | 20m |
Claim: engineering moves up a layer — an agentic system called Relocation Scout is built with modular skills and sub-agents, structured outputs written into… | ||
| Stop Writing Tone Instructions. Layer Them. | Isadora Martin-Dye | 21m |
A production-tested architecture breaks voice into four executable layers: immutable identity (rules like disclose you are an AI in first response and never… | ||
| Agents in Production: How OpenGov Built and Scaled OG Assist | Gabe De Mesa | 19m |
OpenGov describes a production AI assistant called OG Assist that runs across thousands of state and local governments, built on an Effect-native agent loop in… | ||
| Building an Autonomous Engineering Org | Angie Jones | 18m |
Claims a full conversion of an enterprise engineering org into an agentic one using internal agents like Goose and Claude Code plus Anthropic MCP design… | ||
| Agents Building Agents | Alfonso Graziano | 30m |
Claims a systems workflow: a coding agent (Claude Code) edits the product agent, runs self-healing evals against a golden dataset and scorers, and opens PRs or… | ||
| The 100-Tool Agent Is a Trap | Sohail Shaikh, Ankush Rastogi | 28m |
Fat-agent designs load hundreds of tool schemas into every prompt and show concrete failure modes: accuracy falls from ~78% at 10 tools to ~13–13.6% at… | ||
| Using Spec-Driven Development for Production Workflows | Erik Hanchett | 18m |
Claim: spec-driven development splits complex features into three phases — requirements, design, implementation — and uses AI coding assistants to steer work… | ||
| Research to Reality: Bringing Frontier ML Research to Production | Vaidas Razgaitis | 15m |
Uploader promises three tactical tips to speed R&D to production and delivers concrete artifacts and workflows: a research prototype taxonomy document that… | ||
| Structuring the Unstructured | Cedric Clyburn | 21m |
Claims that a local open source pipeline converts PDFs, scans, tables and images into structured JSON/Markdown, demonstrates exporting eight tables from one… | ||
| HTML is All You Need (for Agents to Make Graphics) | Amol Kapoor | 07m |
Claim: agents fail at canvas tools but succeed when given structure. Models like Claude and ChatGPT produce garbage SVGs on Simon Willison’s pelican riding a… | ||
| OpenClaw in Your Hand: Building a Physical AI Terminal | Lech Kalinowski | 25m |
Vault handheld terminal is a text-first dual-display device built on a single ESP32-S3 and one battery cell, pairing a fast emissive OLED with a slow bistable… | ||
| Bypassing the Multimodal Tax: Hybrid RAG, SQL RRF & UI Telemetry | Abed Matini | 46m |
Claims a working local-first RAG prototype that cuts cloud vision token bills by converting docs to Markdown and running OCR locally with local OCR pipelines… | ||
| When All Context Matters: Extended Cache Augmented Generation | Luis Romero-Sevilla | 06m |
Claims a new method called Extended Cache Augmented Generation that shards rapidly changing, fully relevant document collections across multiple parallel CAGs… | ||
| AI System Design: From Idea to Production | Apoorva Joshi | 29m |
Framework for taking an AI idea to production with concrete decisions not platitudes The talk claims a measurable goal to reduce processing time from 2 days to… | ||
| Your Agent Is Wasting Tokens and You Don't Know It | Erik Hanchett | 06m |
Concrete fixes and numbers: code examples add prompt caching (system and tool prompts) so repeated calls send far fewer tokens; model routing by difficulty… | ||
| We Cut 94% of AI Coding Tokens With a Local Code Index | Rajkumar Sakthivel | 11m |
Claims a 94% reduction in tokens by inserting a local retrieval layer between codebase and agents and sending only relevant snippets instead of full files… | ||
| AI-Driven Multi-Document Correlation for Financial Compliance | Varsha Shah | 19m |
Framework combines graph-based entity correlation, adaptive probabilistic risk modeling, and cross-jurisdictional normalization into a deployable system… | ||
| Voice In, Visuals Out: The Agony and the Ecstasy | Allen Pike | 13m |
Claims that voice-in visuals-out is viable now with concrete tradeoffs and engineering fixes and demonstrates an agent that files a Linear issue in real time… | ||
| Your Agent Failed in Prod. Good Luck Reproducing It. | Tisha Chawla, Susheem Koul | 14m |
An architecture for reproducible postmortems that records every model call, tool payload, memory read and state transition into an append-only event log and… | ||
| Using RL Agent to Detect and Remediate ETL Pipeline Failures | Anna Marie Benzon | 15m |
An RL-guided pipeline health agent runs on AWS Lambda + Glue + EventBridge, uses CloudWatch logs and Glue schema metadata, and combines deterministic anomaly… | ||
| You Can't Prompt the Room: The Last Skill AI Won't Replace | Balázs Horváth | 16m |
A hackathon generates 21 ideas, 17 abandoned, 4 survived because of lacking data access, no clear business owner, or no measurable value; the four that survive… | ||
| Deterministic Infra for Non-Deterministic AI Agents | Nishant Gupta | 07m |
Describes a deployed agentic control plane that enforces separation: the model proposes, a policy engine approves, an execution gateway enforces, with audits… | ||
| The Agentic AI Engineer | Benedikt Sanftl | 35m |
Claims an Eval-Driven Developement Loop that automates spec → build → evaluate → diagnose → monitor → optimise for agents, replacing slow human review with a… | ||
| The Prompt is the Platform | Dominik Tornow | 18m |
Claim: the prompt is the platform and the specification becomes the product. Agents initially fail jumping from abstract spec to a Postgres Rust server because… | ||
| User Signal Dies at the Retrieval Boundary | Sonam Pankaj | 16m |
Claims a concrete fix: utility-ranked memory that re-ranks retrieval by outcome history, not just semantic similarity, raising helpful memories and lowering… | ||
| Browser Agents Don't Need Better Models. They Need Better Eyes. | Kushan Raj | 04m |
Claim: the bottleneck is the browser interface not the model. A runtime uses a compact page representation, fast actions with stable handles, and step-by-step… | ||
| WF2026: Software Factories & Keynotes ft. Microsoft, OpenAI, OpenClaw, Z.ai (GLM), MiniMax, HF | OpenClaw, Z.ai (GLM), MiniMax | 08h 37m |
Skip the fluff unless the talk names what was built, how it’s measured, and the actual results. This stream delivers that: Microsoft demos Foundry IQ plus an… | ||
| Chat and citations won't save your vertical AI | Atul Ramachandran | 15m |
Filed builds AI data entry for tax firms and hits 80%+ vs 50–60% baseline yet users still audit every result; the team discovers chat and citation UIs push… | ||
| The Missing Layer After Launch | Raphael Kalandadze | 20m |
A production system runs agents that watch thousands of live conversations and close the loop: a log-monitoring agent scans hourly, deep‑dives traces and opens… | ||
| What Does Done Even Mean? Agents and Paperclip's Liveness Model | Dotta | 07m |
Gives a concrete liveness model: a control plane that enforces three invariants (keep productive work moving, only real blockers stop progress, bound infinite… | ||
| Enterprise Agents Have a Structure Problem | Ishita Daga | 12m |
The paper claims most enterprise agents fail because models lack a machine‑readable metadata layer and instead guess at raw schemas and KBs; the proposed fix… | ||
| Your Voice Agent Doesn't Need a Frontier Model | Joel Allou, Ornella Bahidika | 06m |
Ace runs on Haiku 4.5 and trades model thinking for engineered orchestration: lesson planning, memory of student state, and next‑step decisions move into a… | ||
| Don't Let the LLM Drive | Ornella Bahidika, Joel Allou | 06m |
Ace is a production voice tutor that runs full lessons reliably by moving control out of the LLM into a harness that enforces a small state machine and… | ||
| I Run a Fleet of AI Agents Across Three Machines. Here's What Broke. | Kyle Jaejun Lee | 09m |
Runs a fleet of AI agents across three machines and reports concrete failures and fixes: a hierarchical CEO VP manager worker design that reduces human… | ||
| Why Your Agent Disagrees With Itself (And What To Do About It) | Diane Lin | 26m |
Disagreement concentrates in a decision gray zone not a broken model; using active learning to surface examples, then resolving them with semantic memory… | ||
| Develop at Idea Velocity | Jeffrey Lee-Chan | 15m |
Claim: the bottleneck is the harness not the model — after 1,000 hours of running fleets, agents pass on first runs then fail without persistent memory and… | ||
| Your LLM Deception Monitor Is Broken. The Fix Is in the Training Data | Sachin Kumar | 14m |
A concrete fix detects directional shifts in activations by subtracting base from fine‑tuned checkpoints, training a sparse autoencoder on that delta and… | ||
| ReviewDebt: a practical framework for scoring every pull request | Sachin Gupta | 25m |
ReviewDebt introduces a deterministic score that computes review burden from five signal families — diff size and coupling, test-evidence delta, ownership… | ||
| Stop AI Agent Hallucinations: 5 Techniques + Production Patterns | Elizabeth Fuentes Leone | 55m |
Hands‑on code changes that replace prompt hacks with measurable fixes: Graph‑RAG (Neo4j): 73% fewer fabricated statistics via Cypher queries instead of vector… | ||
| MCP Apps: Primitives, discovery, and the Future of Software | Pietro Zullo | 29m |
Claims that MCP Apps are a bidirectional, stateful UI layer backed by concrete primitives: ui/update-model-context, ui/message, App Tools. Architecture uses… | ||
| Every Solo Agent Builder Eventually Reinvents a Worse Version of CI/CD | Sumaiya Shrabony | 11m |
A real open‑source agentic-content-system — a 19‑skill Claude Code pipeline that runs every other Saturday, reads a knowledge vault, and produces 12 content… | ||
| Semantic Blindness: 500,000 Sensors Confused an LLM | Raahul Singh, Vanč Levstik | 16m |
Claims a failure mode called Topology Trap and two others — sensors-as-strings confuse embeddings, massive context causes attention collapse, and… | ||
| SWE-Marathon: Evaluating Coding Agents at Billion-Token Scale | Rishi Desai | 13m |
SWE-Marathon benchmarks project-scale autonomous coding across 20 tasks and claims to measure coherence at billion-token budgets. It runs multi-hour agent… | ||
| The Prompt Is Still a Punch Card | Ted Johnson | 20m |
The central claim is that prompting still uses a prompt is batch like punch card protocol: assemble a full turn, submit, wait, repeat, even though model… | ||
| The AI bugpocalypse is here. Now what? | Jack Cable | 20m |
Frontier models now find real vulnerabilities in production libraries at scale, not toys, with examples like Mythos and Fable and documented exploits such as… | ||
| A Song of Types and Agents | Roberto Stagi | 14m |
Claim: the application and agent layer shifts from Python to TypeScript, evidenced by GitHub topping in Aug 2025, Anthropic acquiring Bun, and coding agents… | ||
| The Future Is Domain-Specific Agents | Justin Schroeder | 31m |
Argues a composition architecture where many small domain-specific agents (Gmail, Sheets, Notion, GitHub) talk via plain English to a coordinator, each with… | ||
| Frontier results, on device | Rachel Nabors | 31m |
A concrete pipeline uses Phoenix (Arize) to run capability and production-trace evals comparing Claude Sonnet, Gemma 4 E2B, Qwen 2.5 and Llama 3.2 3B against a… | ||
| Medic for Apache Spark - First Aid for Failing Jobs | Drasko Profirovic | 11m |
A diagnostics agent called Medic for Apache Spark automatically ingests logs and metadata and returns evidence-backed root causes and fixes in minutes rather… | ||
| Skills are the New SDKs | Elvin Aghammadzada | 27m |
The team claims that enterprise agent teachability requires versioned, task-specific packages that encode workflow ordering, failure modes and security… | ||
| Designing Voice Agents for Real Conversations | Chintan Agrawal, Daniel Wirjo | 33m |
Production voice agent code using Pipecat plus Daily (WebRTC) and any STT/TTS (Deepgram AssemblyAI Whisper) demonstrates concrete fixes: streaming pipeline… | ||
| Continual Learning for AI Agents: From Failures to Durable Improvements | Soheil Feizi | 23m |
Claims a working system at rely.ai that converts session logs and feedback into replayable learning environments, routes fixes to memory, harness, or model… | ||
| GTM Is You | Victoria Melnikova | 13m |
Evil Martians' product market fit compass claims a formula from analysis of 37 devtools and finds 9 out of 10 startups with product signal stronger than… | ||
| Agent Output Is Not UX: Rendering Layer Your LLM Pipeline Is Missing | Bala Ramdoss | 14m |
Argues that the missing engineering layer is a typed, versioned rendering contract that enforces a fallback for unknown content types so new model outputs… | ||
| The Pipeline Is Dead | Iris ten Teije | 20m |
Claim: the one-version pipeline is obsolete because code can be produced at runtime and personalized for each user. Built prototypes at Differ that deploy a… | ||
| Your Agents Need a Save Button | Hamza Tahir | 17m |
Argues that adding a save button — durable checkpoints of code state, files, tool calls and artifacts — lets teams freeze agents, drop compute to zero, then… | ||
| Your AI Product Will Fail Unless You Can Explain It | Veronica Hylak | 06m |
Claims that matter: a product that cuts a 30 minute support search to search it and get the answer in 10 seconds, an agent observability layer that stops a… | ||
| Build the AI GTM Agent That Knows the Buyer | Sajjan Kanukolanu | 26m |
Claims an AI-first GTM architecture that knows buyers before the first message, built across multiple brands and 75+ AI agents with 18+ vertical knowledge… | ||
| The UX of AI: Making AI-Powered Apps Your Users Don't Hate | Kathryn Grayson Nanz | 36m |
If wanting actionable UX patterns and concrete UI mechanics, watch: the talk lists five specific user problems—trust, clarity, control, transparency… | ||
| Can Oncology Workflows Run Without Human Touch? | Anant Shankhdhar | 17m |
Claims end-to-end automation of oncology prior authorization using an LLM-driven stack: an ingestion agent, EV agent, medical necessity agent and a submission… | ||
| When Agents Meet Physical Data: The Other Physics of Agent Harnesses | Dmitry Petrov | 28m |
Claim: recompute is the enemy for video and sensor data — materialize once, recall cheaply. Demonstrations use Claude Code over raw S3 with a YOLO pass that… | ||
| Respect The Process | Andrew Dumit | 17m |
A domain-specific harness constrains effects not expression: it exposes a TypeScript SDK as the only edit door, enforces linting and conflict checks, runs a… | ||
| 500 people vibe-coded for 30 days. I was one of them. | Sanja Grbic | 18m |
Automattic pauses the roadmap for a 30‑day Radical Speed Month where about 500 people launch 794 projects and one participant ships three products. Teams pair… | ||
| Build AI Systems for Discernment, Not Approval | Angel Ortmann Lee | 26m |
Wharton study Shaw & Nave finds people accept AI answers over 80% of the time; Duolingo builds a keystroke copy‑typing detector with a 1% false positive model… | ||
| What if the harness mattered more than the model? | Aditya Bhargava | 32m |
Argues that a better harness can make weaker local models competitive, citing Harness Bench numbers that jump from 52.4% to 76.2% with only harness changes… | ||
| How we taught agents to use good retrieval | Hanna Lichtenberg | 14m |
The talk claims Mixedbread closes the Oracle gap by teaching agents to choose among four real tools — overview search, semantic search, filter chunks and grep… | ||
| Beyond the Harness: A Journey Towards Adaptative Engineering | Rajiv Chandegra | 37m |
Claim: engineering must shift from prewired harnesses to adaptive harness emerges mid-run that self-organizes via agent interactions. Specifics offered… | ||
| The Factory That Dreams: 39 AI Agents, No Framework | Rushabh Doshi | 10m |
A factory-grade company brain runs core ops — sales, quoting, recruiting, marketing, production context — as a multi-agent OS with 39 agents and Athena… | ||
| Building Great Agent Skills: The Missing Manual | Matt Pocock | 21m |
Gives a concrete checklist: pick user-invoked or model-invoked and accept the tradeoff between context tokens versus pilot cognitive load, split skills into… | ||
| WF2026: Autoresearch & Keynotes ft. Anthropic, Google DeepMind, Amazon AGI, Sonar, Arena, Recursive | Anthropic, Google DeepMind, Amazon Agi, Sonar, Arena | 08h 52m |
Roll the tape if concrete claims matter: Anthropic pitches the new Fable mythos class of models with code‑execution overhangs not vague demos; Sonar shows a… | ||
| Agents Need Feature Flags | Sachin Gupta | 19m |
Agents ship six concrete flag types: prompt variant flags, per-tool access flags, model-routing flags, memory-policy flags, autonomy-level flags (suggest vs… | ||
| You Didn't Ship a Bug. You Just Wrote It for a Human. | Ravi Madabhushi | 13m |
A demo agent that calls Gmail Calendar and a handful of connectors runs on a 15‑minute schedule and repeatedly strains the production DB; latency spikes trace… | ||
| Stop Renting Your Cognitive Infrastructure | Thiyagarajan Maruthavanan | 08m |
A lab sees inference costs balloon to hundreds of thousands of dollars and a retailer spends ~$200M while an Uber team burns a year of token budget in four… | ||
| Stop Burning Tokens: Why self-improvement needs domain expertise first | Annabell Schäfer | 18m |
A minimal self‑optimization loop on arXiv paper classification runs on Langfuse with a small base LLM as classifier and a frontier optimizer proposing prompt… | ||
| From Tokens to Cells: Foundation Models for Single-Cell Biology | Akram Baharlouei | 17m |
Single-cell datasets scale to tens to hundreds of millions of cells but are noisy, batchy and inherently heterogeneous, so masked-transformer models that treat… | ||
| From Blind Spots to Merged PRs: Continuous Agentic Performance Optimization | May Walter | 23m |
Claims a production system that runs weekly agentic analyses and surfaces verified fixes that turn into merged PRs, not prototypes; shows P90 latency… | ||
| Build Evals That Actually Matter | Nick Ung, Akshay Sharma | 38m |
Claims a simple cause and shows concrete fixes: offline evals hit 90% offline pass yet production fails because the synthetic user is too polite, so Lyft… | ||
| Autonomous Agents for Scientific Tasks | Sina Shahandeh | 19m |
Claim: autonomous agents improve on real scientific measurement tasks when they generate and test explicit hypotheses. Demonstration: an insilicopet CT→PET… | ||
| Agents Need Receipts, Not More Tool Calls | Armanas Povilionis | 20m |
An open protocol called Froglet publishes a service, another node discovers and invokes it, and a signed receipt proves the whole interaction; the demo runs a… | ||
| A Practitioner's Guide to Graphs | Tim Ainge | 14m |
Speed-run shows a schema-first extractor that turns recipes into triples then into typed nodes with standardized units using an embedding model for fuzzy… | ||
| Content Is Code | Matt Palmer | 11m |
Claims that content shifts left into code and demonstrates it by recreating a Conductor product tour with React and Remotion, showing a working but slightly… | ||
| RLM: Recursive Language Models for Large Codebases | Shashi | 17m |
Claims that a repo loads into a programmable REPL where the model writes inspection code and recursively delegates subquestions via llm_query; an independent… | ||
| The Agentic Web and the Bazaar Era of AI | Ramesh Raskar | 12m |
Claims a concrete stack: an open Nanda index that returns an agent card and a signed agent facts record, a message box that authenticates and queues messages… | ||
| Stop Evaluating Models Like It's the 50s | Alejandro Vidal | 24m |
Claim: applying Item Response Theory to model evals yields per-item difficulty (B), discrimination (A) and per-model theta with real error bars instead of raw… | ||
| AI Agents for Performance: Ship Faster, Pay Less | Rajat Shah | 34m |
An agent reads a profiler call stack, traces a hot path to the exact source line, and turns an quadratic-time tensor merge into an O(N) fix that is validated… | ||
| nothing matches | ||