Skip to content
AI Engineer NYC 2026

What to take from both days. What to try next.

An engineering field guide to AI Engineer NYC 2026, October 13–14. Compare approaches across both days, understand their limits, and choose an experiment for your workload.
Find your field guideBrowse both days

Engineering field guide · October 13–14, 2026 · New York

Jump to practical takeaways ↓

The lessons, connected · Both days

From model output to useful work

Define the task. Give the agent context and controlled tools. Verify the work, then use what you learn to improve the next run.

From model output to useful workA real task moves through relevant context, an agent with tools, and verification to an accepted, useful result. Permissions, isolation, and policy constrain execution. Failures and corrections improve context, skills, and evaluations for the next run. Measure cost per accepted result, latency, and human review time. Each stage links to a field guide. AI ENGINEER NYC 2026 · BOTH DAYSFrom model output to useful work Permissions + isolation Enforce policy in code 01Real taskDefine success first02ContextRelevant · currentLinked to sources03Agent + toolsState · recovery04Verify workFacts · behaviorQuality · sources05Useful resultAccepted · usable Improve the next runContext · skills · evaluations MEASURE THE OUTCOMECost per accepted result · latency · human review timeFrom model output to useful workA real task moves through relevant context, an agent with tools, and verification to an accepted, useful result. Permissions, isolation, and policy constrain execution. Failures and corrections improve context, skills, and evaluations for the next run. Measure cost per accepted result, latency, and human review time. Each stage links to a field guide. FROM MODEL OUTPUT TO USEFUL WORK 01Real taskDefine success first02ContextRelevant · currentLinked to sources03Agent + toolsState · recovery04Verify workFacts · behaviorQuality · sources05Useful resultAccepted · usable Permissions + isolation Policy in code Improve the next runContext · skills · evaluations Cost per accepted resultLatency + human review time
Our synthesis of both days. Follow a stage to read the evidence and try an experiment. The arrows describe an engineering process, not measured correlations.

Explore the supporting topics →Download the diagram (SVG) ↓

Find a lesson for what you are building

Section titled “Find a lesson for what you are building”

Each guide compares approaches from both days, explains where the evidence stops, and suggests an experiment you can measure.

Follow a topic across both days.

Explore 47 distinct talks. See where cost, security, evaluation, and other topics recur. Compare the days and check timestamped source passages.

Explore topics and their evidence →
Want to study an individual talk?

The official talk library ↗ has detailed reading editions, slides, and searchable transcripts for October 13. Our guides connect the engineering lessons across both days. Matching session notes link directly to the official editions.

Practical takeaways for AI engineers

Section titled “Practical takeaways for AI engineers”

Twelve actions drawn from both days. Follow the paired sessions to compare the evidence and trade-offs.

  • Build a test set from your own work. Start with real questions, past tasks, and known failures. Keep the examples as regression tests when you change models or agent context.

    Bridgewater · Day 1 · Blackstone / Norm.ai · Day 2

  • Check the decisions inside long agent runs. Define expected behavior before changing prompts or tools. Review traces against that spec and replay from the decision that went wrong.

    Arca · Day 1 · Basis · Day 2

  • Enforce permissions on every tool call. Use task-scoped access, explicit budgets, and checks outside the model. Bloomberg’s incidents extend Day 1’s authorization lesson to network access and tenant isolation.

    Keycard · Day 1 · Bloomberg · Day 2

  • Keep credentials and audit records outside the agent’s control. Use a proxy for approved access. Protect the rules the agent must follow, and remove credentials when the run ends.

    Modal · Day 1 · Google DeepMind · Day 2

  • Separate model judgment from execution rules. Use the model where interpretation is needed. Keep approvals, required gates, and failure handling in code. Increase autonomy only when repeated evaluations justify it.

    Perpetual · Day 1 · Coinbase · Day 2

  • Keep source facts separate from generated explanations. Adapt schemas to the task, but preserve source identifiers. Retrieve and calculate numbers with code. Refuse unsupported facts instead of relying on a final citation check.

    Southbridge · Day 1 · Kepler · Day 2

  • Compare cost at the quality and latency you need. Measure useful output under your workload before buying capacity. Serving software and operator choices change what the same hardware delivers.

    Revolut · Day 1 · SemiAnalysis · Day 2

  • Test a smaller specialist before using a frontier model everywhere. Use task evaluations to route work. Day 1 shows why model rankings vary by task; Day 2 shows Ramp post-training a specialist beneath a frontier orchestrator.

    Vals AI · Day 1 · Ramp · Day 2

  • Capture expert judgment and track when facts change. Have experts explain decisions when they write skills. Date the facts those skills use, preserve their history, and treat missing updates as missing evidence.

    Range · Day 1 · Citadel Securities · Day 2

  • Turn production traces into the next improvement. Trace the agents building and running your system. Use failures and human feedback to curate evaluations. Test routing or training changes against those examples.

    Arize AI · Day 1 · Prime Intellect · Day 2

  • Measure delivered value, not code volume or token spend alone. Set a baseline for the workflow you want to improve. Track accepted work and total delivery cost. More generated code does not prove more useful output.

    Sid Pednekar · Day 1 · Jellyfish · Day 2 · Sonar · Day 2

  • Give coding agents product requirements before they build. Specify save behavior, undo, keyboard access, and semantic HTML. Check the result in a running product. CAIS’s reported cost saving is one experiment, not a general benchmark.

    Railway · Day 1 · CAIS · Day 2

What the two days add up to

Section titled “What the two days add up to”

Capability needs a system around it

Section titled “Capability needs a system around it”

Day 1 shows agents taking on whole workflows at Arca. Day 2 asks how those workflows earn trust: Coinbase ties autonomy to measured quality, and Basis checks behavior throughout the run. The shared lesson is to design the process, the checks, and the fallback together.

Arca · Day 1 · Coinbase · Day 2 · Basis · Day 2

Safety is an engineering boundary

Section titled “Safety is an engineering boundary”

Keycard and Two Sigma make the case for scoped actions and proportional review. Bloomberg’s Day 2 incidents show why prompts and a sandbox alone are insufficient. Identity, network access, and enforcement must remain outside the model’s discretion.

Keycard · Day 1 · Two Sigma · Day 1 · Bloomberg · Day 2

The denominator matters more than the token price

Section titled “The denominator matters more than the token price”

Revolut measures serving cost and latency. SemiAnalysis measures acceptable output from the same hardware. Jellyfish measures the full cost of a merged pull request. Together, they shift the question from cheap inputs to useful work delivered at the required quality.

Revolut · Day 1 · SemiAnalysis · Day 2 · Jellyfish · Day 2

Institutional context is both an asset and a maintenance job

Section titled “Institutional context is both an asset and a maintenance job”

Bridgewater builds a governed document corpus; Range captures planners’ judgment in skills. Citadel Securities adds a missing dimension on Day 2: facts become stale. Useful context needs dates, source evidence, and a way to revise beliefs without rewriting history.

Bridgewater · Day 1 · Range · Day 1 · Citadel Securities · Day 2

Improvement moves from one response to the whole loop

Section titled “Improvement moves from one response to the whole loop”

Arize turns telemetry into fixes. Ramp tests specialist models beneath its orchestrators. Prime Intellect proposes learning from the organization’s own traces and feedback. These are different stages of the same loop: observe failures, create tests, change the system, and measure again.

Arize AI · Day 1 · Ramp · Day 2 · Prime Intellect · Day 2

Useful talks beyond the main agenda

Section titled “Useful talks beyond the main agenda”

Both streams carried recorded talks during scheduled breaks. Start with these connections; the day digests link the rest.

Voices from both days

Section titled “Voices from both days”

“It’s about how do you build the loop to write that feature.”

Aparna Dhinakaran, Arize AI · Day 1 · Session summary · Watch the moment

“The loop is really only as trustworthy as the check.”

Evan Kormos, Coinbase · Day 2 · Session summary · Watch the moment

“Skynet was, after all, a closed source company.”

Dillon Rolnick, Nous Research · Day 1 · Session summary · Watch the moment

“The reality is every fact has a half-life.”

Joao Wenzel, Citadel Securities · Day 2 · Session summary · Watch the moment

Keep exploring

All session summaries → · Day 1 digest · Day 2 digest · Agenda and stream notes · About these notes · Official talk library ↗

Unofficial notes from the October 13–14 main-stage livestreams. Figures and results are reported by speakers. Check the linked sessions for evidence, scope, and recording limits.