The session connects engineering judgment with product thinking: influence what gets built, account for deployment and operations, and preserve options so technical decisions do not unnecessarily constrain the future.

Product thinking means judging what to build, what works in operation, and what can evolve.

Trisha Gee examines these choices at QCon London 2027.

Read the abstract: buff.ly/CHX1x7w

Their picks span agent authorization, context costs, open-source inference, engineering organizations, agent standards, and audio AI.

Each Program Committee member explains what interests them about their two selections.

Which QCon AI New York sessions are the people behind the program most looking forward to?

Hien Luu, Wes Reisz, and Faye Zhang share six picks.

Read their selections in this InfoQ article: buff.ly/KrZCEcm

More sessions like this are lined up for QCon London 2027, April 13-16.

Find out more about QCon London 2027 here: buff.ly/aBnAUeP

#QConLondon #AIAgents #AIInfrastructure

Batch APIs can look attractive for background agent workloads, but batch latency applies to every request.

Meryem Arik, Co-Founder and CEO at Doubleword, on why an agent with 20 to 30 sequential prompts can take 20 to 30 days end to end.

More sessions like this are lined up for QCon London 2027, April 13-16.

Find out more about QCon London 2027 here: buff.ly/U7LjaAc

#QConLondon #AIAgents #AIEngineering

LinkedIn's first approach to agent memory had each application identify and store its own signals. Every new use case needed another bespoke integration.

Karthik Ramgopal, Distinguished Engineer at LinkedIn, on why the memory layer had to become shared infrastructure.

LinkedIn will share server-side guardrails and peer approvals for an operations agent supporting a Kubernetes platform spanning 500,000+ nodes.

The two-day program also provides time between sessions to compare implementation choices.

Pinterest's session covers disaggregated inference, KV-aware routing, and cache offloading for multimodal workloads.

Datadog will examine how it cut cost per AI interaction to one-third while maintaining quality and latency.

How does Pinterest serve vision-language models efficiently?

Salina Wu and Cristian Lopez will discuss their NVIDIA Dynamo stack at QCon AI New York, Dec 15–16.

See the schedule: buff.ly/JysYDJz

More sessions like this are lined up for QCon London 2027, April 13-16.

Find out more about QCon London 2027 here: buff.ly/SPwJhBf

#QConLondon #AIEngineering #CodingAgents

Jeff Smith, CEO & Co-Founder at Neoteny AI, on why a coding agent working from general model knowledge can't know the rules that make a PR mergeable.

LangChain enforces six distinct rules around lazy imports, at different levels of the repository, and very little of that is documented in one place.

The registry makes agents owned, versioned, deployable artifacts.

Governance adds authorization, audit, and kill switches.

Episodic and procedural memory remain separate because they carry different verification and privacy boundaries.

The stack was built in dependency order: a shared model interface, then evals and guardrails, a common representation for canvas and SDK, an agent registry, governance, and a skills hub.

A registry turns agents into owned, versioned artifacts before governance adds controls.

Sahil Vashishat explains BlackRock's build order at QCon AI New York.

Read the abstract: buff.ly/gvlgCiS

@Doubleword_ More sessions like this are lined up for QCon London 2027, April 13-16.

Find out more about QCon London 2027 here: buff.ly/bISxVox

#QConLondon #AIAgents #AIInfrastructure

Agent reasoning loops can keep running without a useful result while the inference bill grows.

Meryem Arik, Co-Founder and CEO @Doubleword_, on a simple production guardrail: set a cost ceiling per task and flag anything that crosses it.

More sessions like this are lined up for QCon London 2027, April 13-16.

Find out more about QCon London 2027 here: buff.ly/qg1LjJc

#QConLondon #AIAgents #AIEngineering