Vivek Trivedy, who publishes as Viv, leads applied research with LangChain’s Labs team. He works on continual learning, agent harnesses and evaluation data, connecting earlier research in visual representation learning with the engineering of agents that can inspect and improve their work.[1][2][3]
Start with How we build evals for Deep Agents →
GitHub 
X Career
Trivedy studied at Temple as an undergraduate and graduate student, completing a PhD on attention and information selection for visual tasks. He also worked for four years as a scientist at AWS, including life-sciences and healthcare AI. His current LangChain research combines harness design with mining and curating traces to improve agents over time.[2][6]
Contributions
Use traces to choose a targeted harness change
In his account of improving Deep Agents, Trivedy describes a team experiment that kept the model fixed while changing prompts, tools and middleware. Execution traces revealed failures that informed verification loops, environment context and other adjustments. The article explains an experimental process around a particular coding benchmark, rather than attributing every improvement to a different model.[5]
Make each evaluation express a behavior that matters
With Mason Daugherty, Eugene Yurtsev and Harrison Chase, Trivedy coauthored the Deep Agents evaluation guide. It connects datasets, metrics and targeted experiments to behaviors such as reading across files and composing tool calls. The guide warns that accumulating tests can create an illusion of progress when the suite no longer represents production needs. Traces help identify cases, but the evaluation’s purpose still has to be chosen carefully.[3]
Ideas
His Anatomy of an Agent Harness essay defines the harness as the surrounding code, configuration and execution logic. Durable storage, tool access, environment setup and control flow provide abilities that a raw model call does not supply. He works backward from a desired behavior to the machinery needed to support it, giving developers a way to reason about the system’s boundaries.[4]
Tools, writing & talks
Writing, research & talks
Start here
How we build evals for Deep Agentsblog.langchain.comA guide to sourcing data and measuring specific behaviors in an agent harness.
Begin with the examples of behaviors worth measuring, then follow the dataset and metric choices. Basic familiarity with agents helps; the guide’s method can be understood without running Deep Agents or configuring LangSmith.
Coauthor with Mason Daugherty, Eugene Yurtsev and Harrison Chase[3]An X post walking through how to evaluate deep agents — the author's approach to curating eval data, defining metrics, and running experiments to measure whether agent changes actually help. A write-up of an approach, not a package you install.
Catalog workOriginal source- The Anatomy of an Agent Harnessvtrivedy.com
A framework for connecting desired agent behaviors with infrastructure and control flow.
Author[4] - Improving Deep Agents with Harness Engineeringvtrivedy.com
A trace-driven account of changing a coding agent’s harness while keeping its model fixed.
Author; team experiment[5]