QCon AI New York 2026 to discuss agent identity, context cost, and open source inference

Agent identity: who authorizes what an agent does on its own

The QCon AI New York 2026 Program Committee announced, on October 29, the six sessions it considers most relevant among those scheduled for the event, which takes place between December 15 and 16 in Jersey City, United States, according to an InfoQ report. The choices were made by three members of the committee: Hien Luu, conference chair, Wes Reisz, Technical Principal at Thoughtworks and creator of the InfoQ Podcast, and Faye Zhang, Staff Engineer at Google and responsible for autonomous agents at Gemini Enterprise.

Reisz's first pick was the keynote "When Software Becomes a User: Identity and Authorization for Agents in Production," by Nancy Wang, CTO of 1Password. The talk addresses a problem that already shows up in production: delegating authority across chains of subagents and how an agent accesses credentials without exposing secrets in its own context.

As agents take on more autonomous work, traditional identity and authorization models start to break down. Nancy will explore delegated authority, credential management, and securing agent interactions, foundational capabilities I believe we must get right to truly adopt agentic delivery.

Wes Reisz, Technical Principal at Thoughtworks

Context cost: token price doesn't tell the whole story

Reisz's second pick was "Make Each Token Count: Compression, Caching, and Retrieval for Production LLM Context," presented by Tejas Chopra, CEO and co-founder of Headroom Labs. The session covers learned prompt compression, context alignment with prefix caching, and compress-cache-retrieve pipelines, including the failure modes that affect tool calls and agent trajectories.

Token costs and latency are real constraints when building production agents. Tejas will explore compression, caching, and retrieval strategies, particularly token compression, which I believe will become a must-have capability in modern agentic systems.

Wes Reisz, Technical Principal at Thoughtworks

This theme connects with another committee pick: the session by Francisco Javier Arceo, Forward Deployed Engineer at Baseten and maintainer of the vLLM Agentic API project, will trace a coding task from the harness through tool execution, passing through model generation with vLLM and llm-d. According to InfoQ's summary, the session discusses where tools run, who owns the context, and why cost per completed task matters more than token price alone.

What changes in the engineering organization

Hien Luu's pick was "The Future of the Engineering Org," by Peter Bell, founder and CTO of Gather.dev. The talk draws on conversations with engineering leaders at companies like Anthropic, OpenAI, Shopify, and Coinbase to discuss how coding agents are changing team size, rituals, hiring, and engineer evaluation, and the future of junior developers, a topic that gains urgency as agents take on tasks that used to serve as training for people starting their careers.

Standards for agents and the future of audio with LLMs

Faye Zhang described her two picks as sessions "at the intersection of frontier research and production engineering." The first is "Navigating Agentic Standards: What to Build On Now and What to Watch," by Kevin Kelley, Technical Lead for Standards at Anthropic, which organizes the landscape of standards like MCP, A2A, and OAuth around questions such as which agent is acting, on whose behalf, and how it reaches tools and other agents, separating what is already settled from what remains open.

Zhang's second pick is the keynote "Audio Processing with Large Language Models," by Tara Sainath, Distinguished Research Scientist at Google DeepMind and head of audio for Gemini. The session discusses unified multilingual representations, end-to-end versus cascaded architectures, and the latency-quality trade-offs required by real-time conversational intelligence.

What this means for developers in Brazil

Brazilian teams that have already put agents into production recognize the first two topics in practice. Authorizing an agent that calls tools, reads internal spreadsheets, or triggers actions on behalf of a human user is an architecture problem that still has no standardized solution: most teams solve this today with shared API keys and generic service accounts, exactly the pattern that Nancy Wang's keynote describes as unsustainable at scale.

Context cost weighs even more for those operating in reais, Brazil's currency: token price is charged in dollars, and every redundant tool call or poorly cached context becomes a direct cost on the foreign-currency invoice. The discussion on compression and caching raised by Tejas Chopra and Francisco Arceo's cost-per-completed-task metric give a more precise vocabulary than simply "spending fewer tokens" for teams that already track this expense.

The session on open source inference with vLLM is also relevant to anyone evaluating running their own models instead of relying solely on third-party APIs, an option that is growing among Brazilian companies concerned about cost and data sovereignty. And the conversation about standards like MCP, A2A, and OAuth is relevant to any team that has already integrated MCP into an agent and felt the lack of a clear identity and authorization model between different agents.

Event Details

QCon AI New York 2026 takes place on December 15 and 16 at the Westin Jersey City Newport, one PATH stop from Lower Manhattan. The early-bird price of $1,705 for the two-day pass is valid until November 10; after that, the price rises to $1,790. The full program is available on the conference website, along with an email template for anyone who needs to convince their manager to approve the registration.