Keynotes and Breakouts

What It Was

June 30 was the main programming day for AI Engineer World's Fair 2026 at Moscone West, with the official Autoresearch and Software Factories livestreams acting as the visible spine for a much wider track map. The keynote block established the day's center of gravity: Thariq Shihipar opened with Fable as a field guide for the conference, Tariq Shaukat argued from Sonar's software-quality vantage point that verifiers are becoming the core discipline for agentic software, Antje Barth framed perception agents around human-agent collaboration, Benoit Schillings connected Google DeepMind research to deployed systems, and Laurie Voss with Aparna Dhinakaran positioned evals as a first-class engineering track rather than an afterthought.

The connected session pages make the day read as a live inventory of AI engineering bottlenecks. Autoresearch sessions ranged from Richard Socher's first steps toward automated AI research to Han Xiao's test-time compute for frozen dense-retrieval embeddings, then later toward autonomous research loops, self-improving context harnesses, and AI agents that became serious contributors in software-engineering challenges. The Software Factories and sandbox thread kept returning to operating substrate: Adam Azzam's argument to build environments instead of just agents, Shashank Goyal's adoption story about letting intern-style agents loose, Ivan Burazin's warning that Kubernetes is not an agent sandbox, Samuel Colvin's sandbox-not-desert framing, and Abhishek Bhardwaj's fork-to-fleet sandbox-cloud sessions all point to the same claim that agentic development needs controlled execution environments, durable runtimes, observability, and deployment guardrails before it can be trusted in production.

Several parallel tracks explored what production AI systems are supposed to act on. Computer-use talks asked whether agents should use the web directly rather than waiting for APIs, how close computer-use agents are to statistical failure edges, and what changes when websites become environments for humanlike web automation, multi-cursor action, and context-as-a-service. Context and design-engineering sessions treated context as infrastructure: build-time developer tools failing at runtime, Recursive Language Models, LinkedIn-style codebase skills, context graphs, MCP apps, API-readable catalogs, spatial canvases, and the possibility that chat is the wrong interface for agent work. Evals sessions made measurement concrete through Vending-Bench's simulated business, signal-to-PR self-improvement loops, Uber-scale multimodal closed-loop evals, agent-trace simulations, behavior-shaping prompt/eval loops, and video-slop evaluation.

The day also pushed beyond code into robotics, memory, posttraining, enterprise adoption, governance, and economics. Robotics and world-model talks covered simulation infrastructure, commercial robots, mass-produced humanoids, frontier robotics, drones, and edge robotics. Memory and continual-learning sessions examined static-intelligence limits, long-running research-agent memory harnesses, compute-on-context, gradient-free continual learning, enterprise continual learning, and the idea that improving agents may be a data-mining problem. Enterprise and expo programming grounded those ideas in billing engines, AI spend-to-value tokenomics, governance trends, voice-enabled support agents, air-gapped consumer-data systems, trusted throughput, policy APIs, code-review redesign, mergeable-code context engines, agent-readable catalogs, latency budgets, and weekly shipping systems such as VS Code. This page preserves the official schedule and media evidence layer while linking outward to session pages as they accumulate transcripts, official descriptions, OCR, resources, and synthesis.

Official Livestream Recordings

Scheduled Sessions