QCon AI Boston 2026 (June 1–2 at Boston University's George Sherman Union) is two weeks away and nearly sold out. The full schedule contains 40+ sessions. Six sessions are highlighted below that anchor on the same pressing challenge: what AI engineering looks like after the demo, when teams have spent two years finding out which parts of the prototype don't hold up.
Keeping ChatGPT Fast in the Agentic Era
Martin Spier, ChatGPT Performance @OpenAI
In this day-one keynote, Martin Spier takes on a common misconception: AI application latency is not simply a GPU problem.
A single user request can pass through client work, conversation loading, context assembly, tokenization, routing, inference, streaming, and observability. Any one of those layers can become the bottleneck.
The second half of the problem is newer. Agentic coding lets teams ship faster, which also means performance regressions can accumulate faster. Martin will cover how OpenAI is moving performance engineering toward agent-operated investigation, with telemetry and tooling that agents can read directly.
Context Engineering at LinkedIn: How We Built an Organizational Context Layer for AI Agents with MCP
Ajay Prakash, Senior Staff Software Engineer @LinkedIn
Coding agents work well out of the box until they have to do real work inside a specific company.
They do not know your services. They do not know your internal frameworks. They do not know which data systems matter, which workflows are standard, or which conventions have built up over years of engineering practice.
Ajay Prakash’s session looks at how LinkedIn approached that problem with CAPT, an MCP-based context layer for AI agents. The architecture matters, but the more useful part may be the organizational deployment story: what happened when LinkedIn tried to roll MCP out across engineering, what did not work first, and how the system evolved. Reported results include 70% faster issue triage and 500+ community-authored skills.
The Agent Harness: Control Planes, Invariants, and Approval Boundaries for Production AI Agents
Vinoth Govindarajan, Member of Technical Staff @OpenAI
Vinoth Govindarajan’s talk has one of the clearest positions in the program: agents may appear autonomous, but reliability comes from the harness around the model.
That harness includes control planes, session state, single-writer execution, throttling, tool boundaries, approval paths, and auditability. These are not model features. They are systems concerns.
The session uses OpenClaw as a case study, but the underlying mental model is portable: events enter the system, state is rehydrated per session, execution is constrained, tools are bounded, and every important action leaves an audit trail.
Building Reusable Evaluation Frameworks for Agentic AI Products
Susan Chang, Principal Data Scientist @Elastic
Susan's team has been running a user-facing AI agent in production for almost two years. That duration matters because most production agent stories in circulation right now are no older than six months.
The talk covers her team's eval methods, the centralized eval framework they built to reuse across other GenAI products, and the feedback loop from evals back into product improvement. Worth attending if you are still figuring out which eval pattern matches which failure mode in your own system.
Building GenAI Platform at DoorDash
Siddharth Kodwani, Tech Lead, AI Infrastructure @DoorDash, and Swaroop Chitlur, Staff Engineer / Engineering Manager, Machine Learning Platform @DoorDash
DoorDash’s platform story begins with a familiar failure mode: every team started rebuilding the same LLM plumbing.
Retry logic. Fallback mechanisms. Cost tracking. Prompt versioning. Batch processing. None of it was the product, but all of it was necessary.
This talk explores how DoorDash consolidated that work into shared platform components, including an LLM Gateway, a Batch Inference platform, and an Agentic Gateway. The interesting question is not only how those systems were built, but when shared infrastructure helps and when it becomes another layer of overhead.
Prompt to Prod: Engineering an Autonomous SDLC at Scale
Andrew Swerdlow, Sr. Director of Software @Roblox
More generated code does not automatically mean faster software delivery.
That is the problem this session takes on. AI coding tools can increase output, but production software still has to be reviewed, migrated, maintained, tested, and shipped safely. Roblox’s approach is to treat the SDLC itself as the system to redesign, with autonomous agents handling codebase migrations and maintenance work.
The session introduces Exemplar Alignment, an approach for grounding agents in expert engineering judgment. The harder question is one many teams will face soon: how do you measure quality when an agent generated the code, but a human still owns the production outcome?
QCon AI Boston 2026 runs June 1–2 at Boston University, George Sherman Union. The full schedule and registration are available at boston.qcon.ai.