What is a company AI harness and why does it matter more than the model?

Answered by Falk Gottlob · Reviewed · 3 min read

THE SHORT ANSWER

A company AI harness is the infrastructure that makes a foundation model useful inside a real organization: agents living where employees work, identity and on-behalf-of permissions, persistent memory, sandboxed code execution, and ambient scheduling. These five components are worth more than model quality in enterprise AI because they are what makes an agent reliable, auditable, and compounding over time. Building Falkster.AI, the harness was 70 percent of the work. The model was last.

Before any agent at Falkster.AI did useful work, I had to solve the same five problems every time.

The first was identity. The agent needed to operate as an authenticated principal in the systems it touched. Getting that wrong did not create bad data. It created data with no traceable author, which is worse than no data.

The second was memory. Without state across runs, the agent was the same agent every time. It could not know what had already been written, what the roster looked like last week, what ideas had already been used. Stateless is not an agent. It is a calculator.

The third was ambient scheduling. The scan needed to run on a schedule, not because a human remembered to press a button. The edge cases here were not edge cases: what happens when the schedule fires and a previous run is still in progress, what happens when the agent finishes but the commit fails.

The model was fourth. The first posts on this site ran on GPT-4o-mini. The model the scan runs on today is different. The harness made that swap near-invisible.

The five components

Harrison Chase, the chief executive of LangChain, named the same list in a post the week LangChain raised $300M at $3B. The five: agents living where employees work, identity and on-behalf-of permissions, persistent memory, sandboxed code execution, ambient scheduling.

His argument: these five are worth more than model quality in enterprise AI. The company that builds the harness first owns the agent layer regardless of which model wins.

Workday's result makes this concrete. They cut agent deployment from five months to under a day using LangSmith. That five months was the time required to build and validate the harness components for each new agent. The model was not what was taking five months.

What to buy and what to build

Three of the five have security and compliance requirements that make them worth buying from a platform that builds them correctly: identity and permissions, audit and governance, sandboxed code execution. The failure modes for these are expensive to discover in production.

Two are worth building yourself: memory and ambient scheduling. The scheduling architecture is where compound value accumulates. An agent that adjusts its own cadence based on what it found last run is a product. An agent running on a fixed cron someone else set is a utility. That distinction is worth owning.

For the full analysis of LangChain's Series C and which components to own: LangChain Raised $300M. The Bet Is Not the Model. It Is the Harness.

SOURCES

THE LONG VERSION

RELATED ANSWERS