Vivek Trivedy leads applied research at LangChain Labs, where he develops systems that help autonomous agents improve through experience. His work on harness engineering, production-trace analysis and continual learning addresses a practical problem: how to turn an agent’s mistakes into better performance.
Trivedy earned a doctorate in computer science from Temple University in 2025, studying representation learning and visual attention. His computer-vision research helped vision transformers concentrate on meaningful objects without adding inference-time overhead. He also spent four years as an AWS scientist, primarily developing healthcare and life-sciences AI, and subsequently worked on a startup building visual-reasoning agents before joining LangChain.
How agents become more capable
- The harness shapes the intelligence. Trivedy treats an agent as a model plus its operating environment: prompts, tools, memory, permissions and verification. His work on Deep Agents increased a coding agent’s Terminal Bench 2.0 score from 52.8 to 66.5 without changing its underlying model, using better environmental context, time awareness, loop detection and build-verify-fix cycles.
- Production traces become learning material. His approach to agent trace mining converts tool calls, failed plans, user feedback and context-compaction problems into evaluation datasets and targeted experiments. LangSmith Engine searches those traces, surfaces failures and prepares results for human review; exceptionally long traces require searchable external representations because they cannot fit inside another model’s context window.
- Evaluations should recreate real failures. Trivedy’s work on automated evaluation engineering uses repository context and production behavior to construct reproducible tests. He emphasizes inspecting evaluators themselves for shortcuts such as irrelevant citations or exposed answers, and giving agents detailed feedback instead of relying solely on pass-fail scores.
- Choose models for the workload and economics. His model-task-harness fit framework starts with capable models to establish feasibility, then tests cheaper open models, improved scaffolding and domain-specific fine-tuning. Once harness improvements plateau, he favors adapting the model and then refining its harness again.
His independent projects include an illustrated-storybook agent, a themed coding interface and a talking-avatar generator.
