InfoQ

How LinkedIn, Netflix & DoorDash Run Agents, Evals & Inference in Production: QCon AI New York 2026

8.5内容质量
How LinkedIn, Netflix & DoorDash Run Agents, Evals & Inference in Production: QCon AI New York 2026

TL;DR · AI 摘要

LinkedIn、Netflix和DoorDash等公司分享了在生产环境中运行AI代理、评估和推理的架构与实践,涵盖权限管理、评估流程和模型服务优化。

核心要点

  • Figma通过预热池将代理配置延迟降低90%,可用性提升至99.95%
  • LinkedIn的Kubernetes操作代理管理500,000+节点,包含速率限制和删除保护等安全措施
  • 评估流程直接影响变更上线,需结合离线评估与生产行为监控

结构提纲

按章节快速跳转。

  1. QCon AI纽约2026会议发布生产AI系统架构与操作决策的完整日程。

  2. 1Password CTO提出代理身份系统需突破传统人类凭证假设。

  3. ›Figma代理实践

    预热池技术使代理配置延迟降低90%,可用性达99.95%。

  4. LinkedIn通过Kubernetes操作代理实现500,000+节点的生产变更控制。

  5. 需建立离线评估到生产行为的反馈闭环以确保变更质量。

思维导图

用一张图看清主题之间的关系。

查看大纲文本(无障碍 / 无 JS 友好)
  • 生产AI系统架构实践
    • 代理系统
      • Figma预热池技术
      • LinkedIn Kubernetes安全措施
    • 评估流程
      • 离线评估与生产行为关联
    • 行业案例
      • 1Password身份系统
      • Google音频LLM

金句 / Highlights

值得收藏与分享的关键句。

#AI生产环境#代理系统#模型评估#Kubernetes
打开原文

How LinkedIn, Netflix & DoorDash Run Agents, Evals & Inference in Production: QCon AI New York 2026 - InfoQ

InfoQ Homepage News How LinkedIn, Netflix & DoorDash Run Agents, Evals & Inference in Production: QCon AI New York 2026

AI, ML & Data Engineering

QCon San Francisco (Nov 16-20): Deep technical sessions. Peer conversations that change how you think.

How LinkedIn, Netflix & DoorDash Run Agents, Evals & Inference in Production: QCon AI New York 2026

Nov 11, 2026 4 min read

by

  • Artenisa Chatziou

#### Follow us on

Youtube

232K Followers

Linkedin

26K Followers

Instagram

New

RSS

19K Readers

X

57.1k Followers

Facebook

21K Likes

Bluesky

Listen to this article -

0:00

Audio ready to play

Your browser does not support the audio element.

Normal

1.25x

1.5x

Like

new dropdown rading list

  • Reading list

QCon AI New York 2026 has published its full schedule for December 15–16. Engineers from LinkedIn, Netflix, DoorDash, Figma, and other organizations will examine the architecture and operational decisions behind production AI systems, including agent authorization, evaluation pipelines, model serving, and AI-assisted software delivery. With the conference a few weeks away, there is still time to join other practitioners before the full ticket price takes effect on December 9.

The program addresses questions facing teams moving beyond initial AI deployments: where to enforce agent permissions, how to connect offline evaluations to production behavior, and how to support growing inference workloads. It also examines how coding agents affect code review, engineering practices, and organizational adoption.

The following presentations explore these challenges through systems the speakers build and operate.

Keynotes

Nancy Wang, CTO at 1Password, who founded and scaled AWS Data Protection and previously worked at Rubrik and Google, opens day one with When Software Becomes a User: Identity and Authorization for Agents in Production . She argues that identity systems assume a human behind every credential, with a stable role, a bounded session, and someone accountable afterward. Agents have none of these.

Tara Sainath, Distinguished Research Scientist at Google DeepMind, lead of the Gemini Audio area and modeling lead for Project Astra, opens day two with Audio Processing with Large Language Models . She covers how LLMs have changed audio processing, and the research and product impact of audio LLMs at Google.

Agents in production

Pratik Agarwal, Distributed Systems Engineer on Figma's sandbox and caching platforms, explains how Figma provisions and secures production agents . Figma found that the trust boundary for agents extends beyond the sandbox. Warm pools cut provisioning latency by 90%, and availability rose from 99% to 99.95%.

Ronak Nathani, Principal Staff Software Engineer in Compute Infrastructure at LinkedIn and host of the Software Misadventures podcast, presents LinkedIn's Kubernetes ops agent , which works across 500,000+ nodes and 5M+ pods. Its guardrails sit on the server, including rate limits, delete protection, and peer approval for production changes.

Evals that decide whether a change ships

Fabio Flores, Software Engineer at DoorDash who leads foundation development for its agentic ordering system, explains the eval loop behind it . Every change is checked before launch by an LLM judge calibrated against human reviewers. The same rubrics score simulated and real sessions, so offline gains carry into production.

Emrick Donadei, Software Engineer on the Google DeepMind infrastructure team who works on evaluation and safety for agentic harnesses, covers agent evals the whole organization can write and keep running . The talk starts from a common problem: an eval that passes once, then fails on the next run.

Inference cost and model serving

Fan Yang, Staff Software Engineer building Netflix's AI serving platform, and Rajat Shah, Staff Software Engineer who leads the technical architecture for ML model serving, explain how Netflix serves ML and GenAI from one platform . It handles about 1 million inference requests per second across 300+ models, after five years of consolidation. They cover what multi-tenancy costs to operate.

Salina Wu, Software Engineer building Pinterest's generative AI serving platform, and Cristian Lopez, Software Engineer on GenAI Serving at Pinterest, present Pinterest's vision-language model serving stack on NVIDIA Dynamo . It is built for multimodal workloads with heavy prefill computation and large KV cache pressure.

AI-assisted software engineering

Min Chen, Principal Staff Engineer at LinkedIn with more than 25 years of software development experience, presents LinkedIn's AI code review system . It reviews tens of thousands of pull requests a week across about 10,000 repositories. More than 60% of sampled suggestions are accepted, and a validation pipeline filters findings before developers see them.

Joey Wang, engineering lead at Harvey who created Spectre, Harvey's internal collaborative cloud agent platform, explains how Harvey shipped its agents internally first on Spectre . Spectre passed 7,000 merged pull requests across every repository before the agents reached lawyers.

Engineering leadership and adoption

Liz Fong-Jones, Technical Fellow at Honeycomb and site reliability engineer with more than two decades of experience, and Rebecca Murphey, Principal for AI Ops at Honeycomb and author of Build: Elements of an Effective Software Organization , reflect on Honeycomb's 2x initiative . Peak weekday PR merges to Honeycomb's main monorepo rose from about 30 to about 74 between early 2025 and April 2026, and have held steady since.

Brian Yang, Member of Technical Staff at OpenAI who builds AI-native advertising and monetization systems, covers making AI adoption work across an engineering team . He looks at what happens after every engineer has coding agents, when one person finds a way of working that teammates struggle to reproduce.

Also on the program

Other sessions come from engineers at Anthropic, Microsoft, BlackRock, Databricks, Datadog, Tessl, LangChain, Baseten, and Google Cloud’s DORA team . Topics include agentic standards, sandbox density, SRE agents, AgentOps for financial services, inference with vLLM, voice agents, software factories, and enterprise AI adoption. Explore the full schedule .

Group registration for teams of three or more closes on November 27, 2026, with savings of $50 to $125 per ticket. Individual tickets are $1,790 until December 8, the last price before the full rate applies. Tickets can be transferred to a colleague free of charge until December 11, 2026.

Engineers who need approval can use the Convince Your Manager page , which provides an email template built around the sessions relevant to their work.

main wrapper for authors section

About the Author

section title

main wrapper for each author

#### Artenisa Chatziou

Show more

Show less

#### This content is in the AI, ML & Data Engineering topic

##### Related Topics:

  • AI, ML & Data Engineering
  • QCon AI New York 2026
  • AI coding agents
  • QCon Software Development Conference
  • Artificial Intelligence
  • Security
  • Retrieval-Augmented Generation
  • Software Engineering
  • Related Editorial
  • Related Sponsors The Practical Agentic Coding Playbook
  • Related Sponsor Automated code transformation at scale. Fix vulnerabilities, migrate frameworks, and modernize legacy code across every repo. Try Moderne Free .

The InfoQ Newsletter

A round-up of last week’s content on InfoQ sent out every Tuesday. Join a community of over 250,000 senior developers. View an example

We protect your privacy.