AI Engineer New York Session Guide: How to Navigate 137 Talks

The AI Engineer New York conference schedule is packed with 137 sessions, with up to five running in parallel at any given time. Trying to attend everything is impossible — you have to make choices. This guide highlights the must-anchor sessions and evening events for each day, helping you spend your conference time where it counts.

Monday: Workshop Day — Come With a Real Problem

The best strategy for Workshop Day is simple: pick the session that maps to a problem sitting on your desk right now, and stay for the whole thing. CJ Avilla and Harrison Stall's Claude Managed Agents Workshop kicks off at 9:00 AM in the Sutton room — a solid choice for teams building managed agents.

If "how do we measure agent quality" is your current headache, head to the Turing room instead. Laurie Voss's Evaluations with Phoenix 101 builds from the ground up on evaluation fundamentals. Evals are the undercurrent running through the entire conference, and this session is the right place to start.

The day's biggest opportunity comes at 1:15 PM in the Sutton room. Nick Nisi and Zack Proser's Beyond "Done": Building Verification Loops for AI Agents tackles the central question hanging over the whole week: how does a team trust the results an agent hands back? Verification loops are the critical piece that moves agents from "runs" to "usable." The Expo floor opens at 4:00 PM, with a welcome reception running from 5:00 to 7:00 PM.

Verification Loop refers to a mechanism that automatically or semi-automatically checks, corrects, and provides feedback on an agent's output after it completes a task. Unlike traditional unit tests in software, agent outputs are often natural language or multi-step action sequences that can't be covered by simple assertions — so teams need to design dedicated "judge models" or rule engines to evaluate whether results meet expectations. Typical implementations include: using another LLM to score outputs, replaying results in a sandbox environment to check for side effects, and hybrid approaches combining manual sampling with CI pipelines. Agent systems without verification loops may look impressive in demos, but errors accumulate silently once deployed to production.

Tuesday: Trust, Money, and Agents That Can Pay

Get there early for Tuesday's keynotes. Max Kanat-Alexander opens at 9:54 with "Does Code Quality Even Matter?", followed immediately by Aparna Dhinakaran at 10:12 asking "What does an AI engineer do?" — both big questions, both wrapped up before 10:30. At 10:40, Freddy Lecue presents Wells Fargo's foundation model case study for large-scale fraud detection, bringing the conversation back to real-world production deployments.

The evaluations track has a particularly strong morning lineup. "Your Financial Agent Passed Every Eval. Can You Defend Its Decision?" at 11:30, followed by "Lies, Damned Lies, and Evals" at 11:55 — both sessions are circling the same core question: when a regulator starts asking questions, how much is a polished benchmark score actually worth?

Anyone focused on agentic payments shouldn't miss Jan Curn's "Agents can pay. So why don't they?" (12:20, Shannon room) and Ishween Kaur's "Your Agent Just Paid for Nothing (6 Attacks on x402 Payments)" (2:55, Turing room). Banks and card networks are currently moving agent payments out of the pilot phase — the offensive and defensive dynamics here are worth watching closely.

For those skewing toward security, the Intent Based Auth roundtable at 11:30 in the Simons room has Dick Hardt, Allie Howe, Ian Livingstone, and Michael Davis on stage together to discuss how you define an agent's identity and what resources it's allowed to touch.

Tuesday evening is busy. Table Stakes, a ping-pong happy hour co-hosted by Arize, Modal, Google DeepMind, and others, runs from 5:30 to 8:30 PM; Keycard, Braintrust, and Socket are moving demos and drinks to Elsie Rooftop; and Optiver's Think Fast! trivia night at M Social goes until 10:00 PM. Note: all of these require advance RSVP.

x402 is an experimental payment protocol built on HTTP's 402 (Payment Required) status code, allowing AI agents to autonomously complete micropayments when calling APIs or accessing paid resources — no human intervention required. The core idea: a server declares the required fee and payment address in response headers; when the agent client receives a 402, it automatically attaches an on-chain payment credential (typically a stablecoin transaction); the server verifies it and fulfills the request. This mechanism gives agents a form of "economic autonomy," but also introduces new attack surfaces — the six attack types Ishween Kaur covers include replay attacks, fee manipulation, and forged payment receipts. Intent-Based Auth is a complementary identity and permissions model that differs from traditional OAuth's application-level authorization: it attempts to declare an agent's specific intent at each call, making permissions granular down to individual operations, thereby reducing the risk of over-authorization.

Wednesday: Can AI Actually Do This Job?

Day two of the main conference gives skeptics plenty of room. At 11:05 in the Shannon room, Ryan Stevens explains why giving a financial agent more context actually makes it perform worse — a direct challenge to the intuition that more context is always better. The McCarthy room pulls the hype straight in front of finance's hardest questions: "Can AI run Goldman Sachs?" at 11:30, followed by "Agents Can Rewrite Compilers. Why Can't They Pick Stocks?" at 12:20. Don't miss Maty Bohacek's "Two Ways to Rob an AI-Native Bank" in the same room at 11:55.

For those thinking about their career trajectory, Addy Osmani and Manish Kapur's "Who verifies the software factory?" (11:30, Sutton room) pairs well with Laurie Voss's "The Rise of the Product Engineer" (12:20, Shannon room) — taken together, they sketch out how the engineer's role is evolving in the age of AI. Runway CEO Cristóbal Valenzuela closes the keynote stage at 5:06 with "Everything will be Simulated" — an appropriately grand title for a closing act.

"More context is always better" is a widely held intuition in LLM application development — since model context windows keep expanding, stuffing in all relevant information seems like a safe bet. But real-world financial applications are challenging this assumption: overly long contexts introduce noise, dilute key signals, and trigger the "lost-in-the-middle" effect, where models pay significantly more attention to information at the beginning and end of a context window than to what's in the middle. For financial agents that require precise reasoning, a carefully filtered short context often outperforms a long one packed with raw data. This is why reranking and context compression steps in retrieval-augmented generation (RAG) pipelines are getting more attention in production systems.

Thursday: The Afterglow

Don't tune out for the final sessions. Thursday evening, Prior Labs and Mistral are hosting The Next Frontier of AI Agents in Midtown; Sonar has an afternoon session at Convene on Madison Avenue on how to build an agentic software factory.

Reminder: Times and rooms are subject to change — always verify against the live on-site schedule before heading anywhere.

Closing Thought: Pick a Through-Line

Looking across the full conference, three threads keep resurfacing: evals, trust verification, and agentic payments. Together, they form the thematic backbone of this year's event. Rather than chasing every hot-sounding title, pick the one thread most relevant to your current work and follow it across the four days — that's how you extract maximum value from 137 sessions.