Sydney Runkle

Sydney Runkle is LangChain’s open-source product manager and a former lead maintainer of Pydantic, the Python data validation library. Her work on Deep Agents shows how she chooses framework defaults. When the team’s evaluations found that a built-in to-do list added cost without a clear gain in task performance, it made the tool optional.

From Pydantic user to lead maintainer

Sydney first used Pydantic to validate API requests and responses while working at small startups. She became a contributor and then interned on the library’s open-source team while finishing her computer science degree at the University of Wisconsin–Madison. In 2024–25, she served as Pydantic’s lead maintainer and as a maintainer of Pydantic AI. PyCon US speaker bio · 2025 Python Software Foundation nominee statement

When a developer asked whether one Python model could serve both a database and an API, Sydney suggested separate aliases for validation and serialization. A model could accept one field name on input and emit another on output. The developer pointed out that FastAPI also used those aliases in its generated OpenAPI schema, which made the suggestion unsuitable for that integration. In a separate exchange, Sydney agreed that developers should be able to configure output aliases on the model itself. Until then, callers had to remember to request them each time they serialized it. Alias discussion · Model configuration issue

Sydney and Victorien Plot documented the new model-level alias controls in Pydantic 2.11. They also explained how the release team reduced memory use by reusing validators and serializers it had already built. That change was especially useful in applications with many models that nested or reused one another. Pydantic 2.11 release

Choosing defaults for agents

Sydney joined LangChain in 2025 as an open-source engineer. That year she was LangGraph’s lead maintainer and a maintainer of LangChain, before moving into open-source product management. She helped ship LangChain and LangGraph 1.0 alongside the open-source team. The LangChain agent started with a simple loop in which the model called tools and used their results to decide what to do next. Developers could add agent middleware to require human approval before a tool call without rewriting the loop. LangChain joining announcement · 2025 maintainer roles · 1.0 release · LangChain product manager announcement

• Make planning earn its place. For Deep Agents 0.7, Sydney and the team tested a planning prompt paired with a to-do tool. Across the models and tasks they evaluated, the feature raised costs without a clear performance gain. Making it opt-in let developers choose when to use it. Deep Agents 0.7
• Measure the cost of finishing the task. In her browser tooling experiments, Sydney compared screenshots with text snapshots that identify controls such as buttons and form fields. Cutting a snapshot off at a fixed length gave the model less text per call, but agents fetched the page again. The cutoff failed to lower the cost per successful task. A more compact representation gave the agent less to read without relying on a fixed cutoff. Browser tooling experiments

Sydney and Nathan Drezner describe how a main agent in Managed Deep Agents can hand a long job to an asynchronous subagent and stay available in the original conversation. If a user spots a mistake, they can send a correction while the work is still underway. Sydney takes that question of timing further in her proposal for proactive agents. Schedules and incoming events can start work in the background. She argues that memory can help an agent judge when taking initiative would be useful to the person it serves. Managed Deep Agents · Proactive Agent Engineering session abstract