Keep pulling the thread on Harrison Chase.
By tuning the harness of LangChain's open-source DeepAgents framework, a researcher improved its ranking on the TerminalBench coding benchmark from 30th to 5th place without changing the underlying model.
The combination of LangChain, LangGraph, and DeepAgents for the agent harness, LangSmith for observability, and Gemini Code Assist for code modification creates the complete feedback loop for a self-improving 'meta harness'.
LangChain's strategic vision for the evolution of LangChain and LangSmith is to create an 'AI, AI engineer'—a system that automates the work of an AI engineer.
Google's Gemini Flash model is a suitable choice for running 'online evals' in a system like LangSmith due to its low cost and high speed.
LangSmith provides an off-the-shelf evaluator that detects when an AI agent is stuck in a loop calling the same tools repeatedly, which indicates confusion or errors.
Google's Gemini models have seen significant adoption and popularity in applications requiring multimodal capabilities.
LangChain is actively researching how agent system prompts and toolsets need to be adapted for different foundation models, such as those from OpenAI, Anthropic, and Google.
The early AI agent project AutoGPT, despite being the fastest-growing GitHub project at its launch in early 2023, is no longer widely used because the underlying language models at the time were not capable enough to support its architecture.
LangChain is observing significant interest in its agent harness framework, DeepAgents, as foundation models have become more capable.
LangChain's DeepAgents framework includes a 'virtual file system' concept that allows knowledge stored in a database to be exposed to an LLM as if it were a standard file system.
The concept of 'memory' for AI agents is a key area of focus for LangChain, but there are currently no industry standards for how agent memory should be structured or implemented.
LangChain views 'memory' as the key technological bridge connecting its open-source agent frameworks with its LangSmith observability platform.