Last updated July 1 at 5:30 PM.
The 2026 AI Engineer World’s Fair brings together the people at the cutting edge of building the models, tools, and infrastructure behind the current agent boom, and I’ll be there for every moment. I’ll update this post throughout the conference with the talks worth knowing about, the patterns emerging across the expo floor, and the ideas that matter beyond any individual announcement.
Day 2: Building AI that works in production
Yesterday’s takeaway was that intelligence is becoming table stakes. Nobody is trying to prove that agents can write code, operate a browser, or complete useful work anymore. The assumption throughout the conference is that the models are already capable, and the interesting engineering problems have shifted toward making that intelligence reliable enough to trust.
Day 2 answered the obvious follow up:
What does that actually look like in practice in the enterprise?
Anthropic released Sonnet 5 (and I reacted to it here) and Fable 5 was reintroduced (my colleague Mateo reacted to that here), so I expected the conference to be dominated by benchmark talk and debates over which frontier lab had taken the lead. Instead, I kept hearing about everything required to make those models useful in production. Even OpenAI is talking about Codex, not ChatGPT 5.6.
1. Reliability is becoming its own engineering discipline
The schedule made the trend impossible to miss. Ameya Bhatawdekar of Braintrust argued “Your Agent Evolved. Your Evals Didn’t.” Vinoth Govindarajan of OpenAI went further with “Your Agent Didn’t Fail. Your Harness Did.” Resolve AI broke down the six pillars of an agent harness designed to resolve production incidents.
That second title stuck with me.
Your agent didn’t fail. Your harness did.
It’s a small shift in language, but an important one. The labs are still racing to build smarter models, and Sonnet 5 and Fable 5 are clear evidence of that. The engineers actually building products are increasingly focused on a different question: how do you build systems that consistently turn that intelligence into reliable software?
2. The hard questions have changed
Very few conversations centered on whether or not an LLM could complete a task. Instead, the questions looked like this: how do you evaluate an agent before it reaches production? How do you debug failures? How do you manage memory without exploding cost? How do you secure agent behavior? How do you improve reliability without simply switching to the newest model?
The focus is shifting from demonstrating capability to engineering dependable systems around that capability.
3. The expo floor was telling the same story
What I loved most about Day 2 was how closely the talks matched what I saw on the expo floor yesterday. LangChain on agent orchestration, Inngest on durable workflows, Vapi on voice agents, CopilotKit on in-app copilots, Arize on evaluation, Sonar on code quality, and Browserbase on computer use.
Each one is solving a problem that only shows up after you’ve already picked a model. It’s the same back office Day 1 pointed to on the expo floor, just visible now from the schedule too.
The competitive advantage is shifting toward the engineering systems that make frontier models safe, reliable, and useful in production. That’s also the problem we’re solving at Arcade.dev: once an agent decides what it wants to do, something still has to determine what it’s allowed to touch.
If Day 1 was about abundant intelligence, Day 2 was about the infrastructure required to put that intelligence to work. I suspect that’s where a large part of AI engineering is headed next.
Day 1: Intelligence is becoming table stakes
Day one was workshop day, and the shift from last year was obvious. Nobody was trying to prove that agents can write code, operate a browser, or do useful work. The sessions assumed the capability and focused on making that work dependable.
1. Trust is becoming an engineering discipline
Laurie Voss of Arize ran “From Vibes to Production: Evaluating and Shipping AI Agents That Work.” Atlassian focused on how orchestration changes the software development lifecycle. Nnenna Ndukwe of Qodo taught teams to build quality gates into agentic coding workflows. Charles Frye of Modal ran “What Is an Inference Engine, Anyway?”, digging into the infrastructure required to turn model capability into something fast, efficient, and usable in production. Accenture and Datadog won best title of the day with “Build a Platform, Unleash an Agent on It… and Watch It Burn!”
Evaluation came roaring back with a vengeance. Braintrust ran an observability workshop, IBM hosted “Evals in AI: A Deep Dive,” and Weights & Biases walked through building an end-to-end agent evaluation pipeline.
When a model only produced text that a human reviewed, evaluation was informal. A developer threw together some tests, eyeballed the results, and decided to ship. That stops working when an agent reviews critical code, updates a customer relationship management system, approves refunds, or changes production systems. At that point, an eval becomes both a measuring stick and a governor.
Think about what the calculator did for accounting. It made arithmetic trivial, so the real work moved to approvals, controls, records, and audits. Last year, we got the calculator. This year, we are building the accounting department.
2. Context engineering has grown beyond retrieval-augmented generation
Reliable context has become its own discipline, and retrieval-augmented generation (RAG) is now one piece of it.
Neo4j argued that “RAG Needs a Map.” Unblocked ran “Beyond RAG: Build a Relational Context Engine from Scratch.” Towards AI focused on compaction, memory, and cost, while Elastic declared that “Vector Isn’t Enough.”
The debate has moved past whether models need external information. The question now is how that information should be represented, connected, compressed, refreshed, secured, retrieved, and paid for.
A basic RAG demo proves that a model can find an answer in a document. A production context system must decide which documents matter, how they relate, who can see them, and which source to trust when two disagree. That is the difference between answering a question and knowing a business well enough to affect its bottom line.
3. The agent economy is building its back office
The expo floor was full of companies selling quality, verification, observability, incident response, code remediation, payments, and production infrastructure. Individually, they looked like separate products. Together, they looked like the departments of an established corporation.
The expo floor was not selling a new genius employee. It was selling the back office required to employ millions of them.
The workshops hinted at one crucial part of that back office: authority. Docker covered the move from approval loops to autonomous agents. PayPal showed how it designs command-line tools for agents. Microsoft focused on protecting hosted agents.
Once an agent decides what should happen, something still has to determine whether it can happen, whose authority it carries, which policies apply, and what gets recorded. That is the action layer, and it is the same problem we work on at Arcade.
Day one made one thing clear: as intelligence becomes more available, the advantage shifts toward the systems that make it useful and trustworthy.
