A New York conference lists real figures from LinkedIn, Netflix and DoorDash on how they limit, test and measure AI helpers that do work on their own.
Some of the biggest tech companies are about to explain, with real numbers, how they run AI helpers at work. The QCon AI New York conference has published its full program for December 15 and 16. Engineers from LinkedIn, Netflix, DoorDash, Figma, Pinterest, Honeycomb, Harvey, Google DeepMind and OpenAI will describe systems they run in production, meaning systems used for real work, not demos.
A pull request is a proposed change to software code. LinkedIn also runs a validation step that filters the AI's findings before developers see them.
Several talks focus on limits. Nancy Wang of 1Password argues that identity systems assume a human behind every login, and agents do not fit that. LinkedIn's agent for managing its computer servers works across 500,000+ nodes. Its guardrails sit on the server and include rate limits, delete protection and peer approval for production changes.
Figma found that the trust boundary for agents extends beyond the sandbox, the walled-off space where an agent runs. It also says keeping spare agent environments ready cut startup delay by 90%, and availability rose from 99% to 99.95%.
DoorDash checks every change to its agentic ordering system before launch. An AI judge, calibrated against human reviewers, scores the results. The same scoring rules apply to simulated and real sessions, so gains in testing carry into real use. A Google DeepMind talk starts from a common problem: a test that passes once, then fails on the next run.
Harvey built an internal agent platform called Spectre and used it first on its own work. It passed 7,000 merged pull requests across every repository before the agents reached lawyers.
Honeycomb says peak weekday code merges to its main codebase rose from about 30 to about 74 between early 2025 and April 2026, and have held steady since. An OpenAI talk looks at what happens when one person finds a good way of working with coding agents and teammates struggle to repeat it.
Individual tickets are $1,790 until December 8. Group registration for teams of three or more closes November 27, with savings of $50 to $125 per ticket. Tickets can be transferred to a colleague free until December 11. The figures above come from the conference's own program descriptions.
Companies running AI helpers at scale limit what agents can do, test every change before launch and track results with numbers.
Mercor, which hires experts to help train AI, raised $750 million at a $20 billion value and says it now earns at a $3 billion yearly pace.
Read more: Mercor, a firm that hires experts to train AI, is now worth $20 billionUNSW, Sydney, La Trobe and Adelaide are putting ChatGPT Edu in students' hands. New graduates will arrive used to working with AI.
Read more: Australian universities roll out ChatGPT to tens of thousands of studentsA Tom's Guide writer asked ChatGPT to rebut the biggest AI doomsday arguments. Its answers favor safeguards and rules over panic, and they show which risks are taken seriously.
Read more: ChatGPT was asked to argue against AI doomsday fears. Here is what it said
