Ankush Gola

Gola completed an electrical engineering bachelor's degree at Princeton in 2015. His undergraduate thesis with Joseph Bolling investigated a binocular stereo-vision system using quadrotor aircraft: two camera viewpoints allow a system to estimate depth. Their code remains on his GitHub profile. Princeton awarded both students its Bradley Dickinson Award for System Design in 2015, recognizing their work designing and implementing complex electronic systems.[1][4][5]

After Princeton, Gola worked at Facebook, where Princeton describes him as a senior software engineer. He left after four years and joined Robust Intelligence, where he met Harrison Chase. LangChain's own history distinguishes Chase's original open-source project in 2022 from the company the two founded in early 2023. Gola now serves as co-founder and CTO. The company's work covers building agents, evaluating their behavior and operating them in production.[1][2][10]

Agent workflows using ordinary functions

Gola's introduction to the LangGraph Functional API shows how developers can add durable execution and human intervention to programs without first drawing an explicit graph of steps. The interface uses entrypoint and task functions, while ordinary loops and conditionals control execution. It shares a runtime with LangGraph's graph-based interface. His examples explain how to save conversation state and stream progress as work proceeds, making the article a practical entry point for programmers bringing an existing application into LangGraph.[6]

Finding useful memories

In his article on semantic search for LangGraph memory, Gola explains how an agent can retrieve information by meaning rather than an exact text match. A query can find a relevant earlier interaction or preference even when its wording differs. The implementation adds similarity search to LangGraph's memory store, alongside filters that restrict the records being considered. This provides a retrieval mechanism for an agent's stored knowledge; it does not by itself verify that a remembered claim is true or decide what the agent should retain.[7]

Storing and searching agent traces

Gola introduced SmithDB in May 2026 as the database behind LangSmith's agent-observability workloads. An agent trace records the model calls, tool actions and other steps within a run. Long-running agents produce nested records that arrive over time, so a debugging interface must find individual steps and reconstruct their relationships. SmithDB combines durable object storage with ingestion and query services. In a follow-up article with Akshay Aurora and Sumedh Arani, Gola describes how the team builds and merges text indexes to make those records searchable. These are accounts of a team-built system, rather than claims that he wrote the database alone.[8][9]

In a 2024 ClickHouse interview, Gola described LangSmith's progression from inspecting agent execution to evaluating changes in prompts and models. A trace helps an engineer locate a failure; evaluation compares behavior across examples to test whether a proposed fix improves it. That distinction gives a newcomer a useful way to read his work: memory and workflow APIs govern what the agent can do, while traces and evaluation help developers understand what it actually did.[11][6][7]