Keep pulling the thread on Harrison Chase.
Harrison Chase strongly believes that providing access to a file system is a mandatory requirement for building effective long-horizon agents.
For AI agents, the source of truth for application behavior is a combination of the code and the execution traces, unlike traditional software where the code is the sole source of truth.
The primary domain where long-horizon AI agents are currently achieving the most success and adoption is software development and coding.
The most effective current applications for long-horizon agents are tasks that produce a "first draft" for human review, such as coding, research, and report generation, due to their lack of perfect reliability.
LangChain provides tools like the Langsmith MCP and a CLI called Langsmith Fetch that enable a coding agent to autonomously pull its own execution traces, diagnose failures, and attempt to fix its own code.
The Langsmith Agent Builder includes a memory feature that allows an agent to modify its own instruction files based on corrective feedback from a user during an interaction.
Klarna uses a long-horizon agent to assist its human customer support staff by running in the background after an issue is escalated to generate a summary report of the interaction.
A common and effective practice in building modern AI agents is to provide them with tools for interacting with a file system, either directly or through a bash interface.
Harrison Chase argues that a significant reason for the popularity of Anthropic's Claude Code is the quality of its agent harness, not just the underlying model.
On the Terminal Bench 2 coding benchmark, which separates performance by model and agent harness, Anthropic's Claude Code is not the top-performing agent.
Harrison Chase believes that OpenAI heavily trains its models on using the Bash shell, making it an effective tool for agents built on OpenAI models.
Harrison Chase believes Anthropic trains its models on explicit file editing tools, and developers can gain performance by leveraging these specific tools when building agents.