QCon AI New York 2026 Sets 23 Sessions on Agent Authorization and Post-Deployment Evaluation

QCon AI New York has locked in a 23-session program for its 2026 edition, with the lineup weighted toward the operational side of AI: how autonomous agents get authorized to act, how production systems are guarded and evaluated once real users touch them, and how teams share inference capacity without burning through their budgets. InfoQ reported the session count and themes. The emphasis marks a shift for a conference track that has spent recent years on model capability and moves the conversation to control, cost, and verification.

Agent authorization is the thread running through several of the confirmed talks. As coding assistants, browser agents, and back-office automation gain the ability to call APIs and move money or data, the hard question is no longer whether the model can perform a task but what permissions it holds, who grants them, and how those grants are revoked. That problem sits at the intersection of identity, policy, and software supply chain security, and it is drawing interest from platform and security engineers who have not traditionally attended AI sessions.

The program also turns to evaluation after deployment, a gap that has frustrated teams running models behind live products. Offline benchmarks rarely predict how a system behaves when prompts drift, retrieval sources change, or users find edge cases the test set never anticipated. Sessions on production guardrails and post-launch evaluation suggest organizers expect practitioners to bring war stories about regressions caught late, silent quality drops, and the instrumentation needed to see them.

Shared inference infrastructure rounds out the confirmed themes. Serving models for many internal teams means deciding who pays for GPU time, how requests are queued and prioritized, and whether to run open-weight models in-house or route to a vendor. Those choices now carry real budget consequences as usage scales, and the conference is treating them as engineering problems rather than procurement decisions. Artenisa Chatziou reported the session details.

QCon has not yet published the full speaker roster and schedule, so the balance between vendor talks and practitioner case studies remains unclear. What to watch is whether the sessions produce reusable patterns for agent permissions and evaluation, or stay at the level of describing problems. For teams already shipping AI features, the more useful outcome would be concrete designs they can adapt, not another survey of risks.

The conference circuit has been flooded with AI content over the past two years, much of it focused on what models can do next. A program built around authorization, guardrails, shared infrastructure, and post-deployment measurement is a bet that the field's pressing questions have moved to the unglamorous work of running these systems safely and affordably at scale.