San Francisco, 18–19 July. The AGI Summit — now in its fifth consecutive year at the Palace of Fine Arts on the north edge of the Presidio — returns this weekend as one of the more informationally dense two-day gatherings on the AI calendar. Roughly 15,000 attendees, more than 200 speakers, and a technical programme that runs from mechanistic interpretability and post-training reasoning research through agentic AI product engineering to the applied deployment stories that are shaping how AI actually gets used inside real businesses.
The 2026 edition arrives at a specific inflection point. The industry has spent the previous eighteen months moving from scaling-first debates to reasoning-first ones, from monolithic model launches to agent-and-tool architectures, and from “look at this demo” energy to production-deployment reality. The AGI Summit’s stage this year reflects that shift. The talks and the panels are dense with the specific engineering, evaluation and governance questions the field is actually working on — not with the futurist framing that dominated the same venue three years ago.
The programme reads like the field’s honest state-of-play
Speaker lineups are the clearest signal of where a technical field’s centre of gravity sits at any moment. The AGI Summit 2026 lineup reads, for anyone who has been watching the AI research and product community, like an unusually candid map of who is doing what.
Aengus Lynch of Anthropic — the first author on the Agentic Misalignment work cited in the Claude 4 system card — is on the alignment and safety track. His research programme, and the broader Anthropic alignment output he sits inside, is one of the two or three most consequential streams of empirical alignment work in the industry. Anyone attending the safety and alignment sessions is doing so partly because the Anthropic team is showing up.
Rohan Varma of OpenAI leads the Codex team and, before that, served as the first product manager at Cursor as the AI-native code editor went from zero to a billion in revenue in a year. His session sits at the intersection of two of the fastest-moving deployment stories in AI: developer tooling (where AI has produced its most measurable enterprise productivity gains) and OpenAI’s coding-agent programme (which is the current benchmark for reasoning-plus-tool-use in production).
From Stanford, Christopher Manning — director of the Stanford AI Lab (SAIL) and co-founder of the Stanford Institute for Human-Centered Artificial Intelligence (HAI) — brings the institutional academic frame. Manning was the mind behind GloVe, the word-embedding work that anchored a decade of natural-language-processing progress. Surya Ganguli, co-inventor of diffusion models, brings the theoretical-foundations frame. And James Zou — whose cardiac AI research earned FDA clearance and the 2025 Overton Prize — brings the applied-science-to-regulated-deployment story that most of the summit’s audience is trying to learn how to run.
Underneath the named speakers, researchers and engineers from Google DeepMind, OpenAI, Anthropic, Microsoft, Recursive and Stripe populate the technical tracks. The industry’s engineering leadership is on the programme in force.
Alignment gets its most operational summit yet
The safety-and-alignment track is likely to be the most substantive it has been at any AGI Summit. The field has moved through the philosophical framing phase (2018–2022), the empirical-research-scaling phase (2022–2024), and into what looks increasingly like an operational phase — where alignment research produces controls, evaluations, and disclosure standards that regulated buyers actually plan procurement around. Anthropic’s Agentic Misalignment research programme, DeepMind’s Frontier Model Forum output, OpenAI’s Preparedness framework work, and the emerging red-teaming benchmarks are all part of this shift.
The AGI Summit’s contribution, for the attending industry audience, is a chance to see the specific empirical work — controlled deception experiments, autonomous replication evaluations, mechanistic interpretability at frontier scale, and identity-and-audit architectures for agent deployments — without the marketing overlay.
Agentic AI at production: the developer-tools case study
The developer-tools deployment story remains the industry’s most measurable production case for agentic AI. Cursor’s zero-to-billion revenue trajectory, Codex’s continuing evolution, GitHub Copilot’s enterprise footprint, and the several serious challengers — Windsurf, Cline, Continue, Aider, Zed AI — collectively represent a category where AI-native workflows have produced sustained enterprise productivity gains that are showing up in engineering-team velocity data.
Rohan Varma’s session will be one of the summit’s most-attended not because coding tools are the industry’s most important application, but because the coding-tools deployment is the case study everybody else — enterprise-workflow builders, healthcare-AI startups, legal-tech companies, financial-services product teams — is trying to reproduce. If AI-native developer tooling scales to industrial velocity, the pattern of adoption tells the rest of enterprise AI what to expect and what to build around.
Diffusion, foundations, and where the science still moves
Surya Ganguli’s session anchors the theoretical-foundations track. Diffusion models — which Ganguli helped originate — remain one of the most important architecture classes in modern AI, and the theoretical understanding of why they work as well as they do is an active research programme. The theoretical-foundations track more broadly is where the field’s next architectural bets get argued about: post-transformer alternatives, mixture-of-experts refinements, memory and long-context architectures, and the interaction between reasoning fine-tunes and base-model capability.
Applied AI in regulated environments: James Zou and the FDA-clearance path
James Zou’s session — anchored on the cardiac-AI programme that earned FDA clearance and the 2025 Overton Prize — will be one of the summit’s most practically useful for the applied-AI-in-healthcare cohort. The regulated-deployment path is the bottleneck most healthcare-AI companies fail at, and the specific playbook — evidence generation, prospective validation, regulatory-strategy sequencing, post-market surveillance — is not one the average AI startup founder has previously navigated. A session led by a researcher who has actually cleared it is high-value programming for the audience.
Compute, sovereignty, and the venture read
The industry-business tracks — compute economics, sovereign AI stacks, model-provenance regulation, workforce implications, and the venture-market read on where the next round of AI capital deploys — round out the programme. Bay AI Circle, which produces the summit, has consistently placed these adjacent-industry sessions alongside the technical work to give the attendee mix a chance to see the applied and the fundamental side by side.
What to watch, on the ground
- The alignment sessions. The Anthropic and DeepMind streams are the field’s most operationally consequential.
- The Codex and Cursor deployment stories. They are the case study the enterprise-AI cohort is trying to reproduce.
- Diffusion and foundations. Where the field’s architectural bets get argued.
- FDA-cleared applied AI. James Zou’s regulated-deployment playbook is high-value.
- Compute and sovereignty. The industrial-policy layer.
Open questions
Does the alignment field converge on shared evaluation standards through the second half of 2026, or does capability-aware release policy remain contested among labs? Does the developer-tools productivity story scale from coding into adjacent knowledge-work verticals within the current cycle, or does that translation take another generation of models? And does sovereign AI procurement — increasingly visible across US, EU, Gulf, and East Asian buyers — actually shift the industry’s revenue mix by 2027, or remain a smaller category next to the hyperscaler build-out?
Bottom line: The AGI Summit 2026 arrives at an unusually honest moment for the field. The infrastructure is being built, the alignment research is producing operational output, the agent-and-tool architectures are shipping to production, and the sovereign-and-regulated-buyer conversation is now a mainstream part of the business track. The Palace of Fine Arts, for two days, will host as concentrated a sample of the industry’s engineering leadership as any single venue on the summer calendar. Agora Media will be reporting on the ground.
