The AI Engineer World’s Fair opens in San Francisco on June 28. It has more than 6,000 attendees, 300 speakers and 29 tracks. That scale matters because this is not a conference about what AI might become. It is a conference about what AI has already become, and what now breaks when companies try to run it at scale.
The people attending are not tourists. They are engineers, founders, product leaders and AI operators who have to ship working systems on Monday morning. Read across the program and one thing becomes clear. The industry has crossed a threshold. The question is no longer whether AI can generate code, answer questions or perform tasks. The question is whether we can verify it, govern it, secure it and trust it enough to let it act.
The biggest bottleneck has moved from generation to verification. AI can now produce more output than humans can review. Uber is running 1,800 AI generated code changes per week. Greptile has analyzed 1 million AI generated pull requests. GitHub Copilot, Factory, Qodo and others are all dealing with the same second order problem. Once machines write faster, humans become the constraint. The next major layer in AI may not be another model company. It may be the verification layer that tells us whether the output is correct, safe and usable.
This is especially clear in software. GitHub CEO Thomas Dohmke is asking what the future of the software development lifecycle looks like. That question alone tells you where we are. Software is being reorganized around AI agents that write code, open pull requests, run tests, respond to failures and iterate. Humans move from authors to supervisors. The companies that figure this out will not just have better tools. They will have a different cost structure.
A new discipline is also emerging around the model. The conference calls it harness engineering. It is the scaffolding, state management, orchestration, memory, tool calls, retries and error recovery that make agents work in production. This matters because many agent failures are not model failures. The model did what it was asked to do. The system around it gave the wrong context, the wrong tool or the wrong recovery path. The model is becoming one part of a larger system. The harness is where a lot of the moat will live.
Prompt engineering is also giving way to context engineering. Prompt engineering asks what instruction to give the model. Context engineering asks what the model should know right now. For long running agents, that is the operating system. Context compaction, memory offloading and cost accounting are becoming core production primitives. A weaker model with excellent context management may beat a stronger model with poor context discipline.
MCP is another signal of where the industry is going. The Model Context Protocol is moving from experiment to infrastructure. Figma, Docker and Microsoft are all building around it. But the same protocol that lets agents connect to tools and data also expands the blast radius. An agent that can read a codebase, call APIs and access internal systems is not a chatbot. It is an actor inside the company. That creates a new security and governance problem.
The same issue appears in regulated industries. Two Sigma is presenting on AI agents that assume employee identity and access internal tools. PayPal is presenting on agent initiated payments across ChatGPT and Google AI Mode. These are serious trust boundaries. What is an agent allowed to do? Whose identity does it act under? Who is liable when it makes a mistake? AI capability has moved faster than enterprise governance. The next 12 to 24 months will be about closing that gap.
Voice is another major surface. Text agents fail quietly. Voice agents fail in public. They interrupt, lag, misunderstand and talk over people. Most voice systems are still pipelines that turn speech into text, send it to an LLM and then turn the answer back into speech. Native speech systems should eventually collapse that stack, but we are still in the difficult gap between possible and reliable.
The open weights ecosystem has also become credible. Hugging Face is hosting millions of models and serving large enterprises. Open models are no longer just research tools. They are becoming part of the production stack. That changes cost, control, privacy and deployment strategy.
The bigger point is this. Coding agents were the proof of concept. Knowledge work agents are the business. Legal analysis, financial research, medical reasoning and investment work will need domain specific tools, workflows and verification systems. You cannot simply point a coding agent at a new domain and expect it to work.
The AI Engineer World’s Fair 2026 is not about AI potential. It is about AI operations. How do you run agents at scale? How do you verify them? How do you govern them? How do you secure them? The generation problem has been substantially solved for a large class of tasks. The operations problem is now the center of gravity. The companies that understand this are building for the next five years. The companies still focused only on generation are solving yesterday’s problem.
The conference runs June 28–July 2, 2026 at Moscone West Convention Center, San Francisco. Please visit https://www.ai.engineer/worldsfair/2026