LangChain Raised $300M. The Bet Is Not the Model. It Is the Harness.

LangChain raised $300M at $3B. Harrison Chase named five components of a company AI harness. I built the same list. The harness was 70 percent of the work.

Falk Gottlob · · 5 min read

Building a Company Falkster cover on navy: five stacked cream component cards labeled with Slack/Teams, Identity, Memory, Sandbox, and Schedule, a teal bracket wrapping all five on the right side, labeled 70%.

LangChain raised $300M at a $3B valuation this week. 7,000 organizations are on LangSmith. Workday cut its agent deployment time from five months to under a day. Clay runs hundreds of millions of agent runs per month on LangSmith.

The number matters. The structure behind it matters more. Harrison Chase, the chief executive of LangChain, named that structure in a separate post the same week.

The short version

  • LangChain's $3B valuation is a bet that the harness layer is worth more than model quality.
  • The harness is the five components that connect a foundation model to real enterprise workflows.
  • Harrison Chase named the five: agents that live where employees work, identity and permissions, persistent memory, sandboxed code execution, and ambient scheduling. I built the same list while building Falkster.AI. The harness was 70 percent of the work. The model was last. Three of the five are worth buying from a platform.
  • One of the five is worth building yourself: ambient scheduling. That is where value builds up over time.

The five components

Chase's post, "The five components of building a company harness," lays out what an organization needs before a model becomes a useful internal agent:

1. Agents that live where employees work. Slack, Teams, email, the systems people already use. An agent accessed through a separate interface gets abandoned. An agent that arrives in the channel where the work happens gets used.

2. Identity and on-behalf-of permissions. The agent needs to act as the right person in downstream systems. When a sales agent creates a Salesforce record, that record should carry the right author. When an engineering agent opens a PR, it should act under the engineer's context, not an anonymous service account. Getting this wrong creates audit gaps that become compliance problems.

3. Memory that persists across runs. A stateless agent resets every time. A session-stateful agent remembers the conversation. A run-persistent agent remembers the last week of activity, the state of its own work, and the context of the organization it operates in. The third is a different product from the second.

4. Sandboxed code execution. An agent that can run code but cannot be contained is a security event waiting to happen. The sandbox is not a convenience; it is a requirement for any agent with write access.

5. Ambient scheduling. The agent runs when the work needs doing, not when a human sends a prompt. This is the component that separates a tool from a system.

The same list from building Falkster.AI

When I built the internal agents that run falkster.com, I did not start from Chase's framework. I arrived at the same list from the problems that blocked progress.

Identity was the first blocker. Before agents could read or write anything meaningful, they needed to operate as an authenticated principal in the systems they touched. Getting that wrong did not just create bad data. It created data with no traceable author, which is worse than no data.

Memory was the second blocker. The first version of the creator scan had no state. Every run was the same run. It could not know what had already been written, what ideas had already been used, what the roster looked like last week. Adding persistent memory was not a feature. It was what made the agent useful rather than repetitive.

Ambient scheduling was the third blocker. The scan needed to run on a schedule, not because a human pressed a button. Getting this right required handling the failure 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? What happens when the model returns something that does not pass the validation gates?

The model was last. The first posts on this site ran on GPT-4o-mini. The model the scan runs on today is a different model entirely. The harness made that swap near-invisible. When the harness is working correctly, the model is an implementation detail.

The harness was 70 percent of the work.

70%

of the work building Falkster.AI was the harness: identity, memory, ambient scheduling. The model came last.

Which to buy, which to build

The $3B valuation is a bet that most companies cannot build all five components themselves. That is correct for three of them.

Identity and permissions, audit and governance, and sandboxed code execution are genuinely hard to build correctly. They have security implications, compliance requirements, and failure modes that are expensive to discover in production. The correct failure mode for identity is "agent cannot act," not "agent acts as the wrong person." Building that correctly requires handling edge cases that only appear at scale. These are worth buying.

Memory and ambient scheduling are worth building yourself when your use case is specific. Memory needs to match how your agents think about state, not how a general platform thinks about it. Ambient scheduling, the component that lets an agent decide when to run next based on what it found, is where compound value accumulates. An agent that adjusts its own cadence based on signal is a product. An agent that runs on a cron schedule someone else set is a utility.

Workday's five-month-to-one-day result is real. It is measuring the time to build and validate the harness components for a new agent, not the time to build the agent itself. Buying the harness platform is what shrank that number. The question is whether Workday's agents, now deployed faster, are compounding faster. That is the number that does not appear in the press release.

The building-a-company hub has the full framework for thinking about which infrastructure to own and which to rent.

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

Sources:

The authority cluster for this post is Building a Company.

Share this post

Frequently asked

What are the five components of a company AI harness?+

According to Harrison Chase at LangChain: agents that live where employees work (Slack, Teams, email), identity and on-behalf-of permissions so the agent acts as the right principal in downstream systems, memory that persists across runs, sandboxed code execution, and ambient scheduling. These are the components that connect a foundation model to real enterprise workflows, independent of which model is used.

What is LangSmith and why did LangChain raise at $3B?+

LangSmith is LangChain's agent engineering platform, used by more than 7,000 organizations as of the Series C announcement. It provides tracing, evaluation, and governance for agents built on any framework or runtime. The $3B valuation reflects LangChain's bet that the harness layer, not the model layer, is where enterprise AI spend will concentrate, because most companies cannot build identity, audit, and sandbox infrastructure themselves.

What does Workday's result with LangSmith mean?+

Workday cut its agent deployment time from five months to under a day using LangSmith's infrastructure. The five-month baseline was the time required to build and validate the harness components for each new agent. That time shrank when the harness was provided as a service. The result is cited in LangChain's Series C announcement.

Which of the five harness components is worth building versus buying?+

Identity and on-behalf-of permissions, audit and governance, and sandboxed code execution are genuinely hard to build correctly and worth buying from a platform. They have security implications, compliance requirements, and failure modes that are expensive to discover in production. Memory and ambient scheduling are worth building yourself when your use case is specific, because the scheduling architecture is where compound value accumulates: an agent that learns from previous runs and adjusts its own cadence is a product, not a utility.

What is Falkster.AI and how does it use these components?+

Falkster.AI builds Heidi, the company brain for AI-native businesses. The internal agents that run falkster.com, including this creator scan, use all five components. Identity: each agent acts under a specific authenticated context. Memory: the creator scan holds state across runs (what has already been written, what creators are on the roster). Ambient scheduling: the scan runs on a schedule without a human prompt. The model is swappable; the harness is what makes the product.

Ship StoriesLangChainLangSmithHarrison Chasecompany harnessAI infrastructurebuild vs buyambient agents
Helpful?

Keep reading

The short answer

PM AI Agent Fleet, Mapped to the 7-Stage Operating SystemA live index of every AI agent for product managers, mapped to the 7 stages of the PM Operating System: Sense, Discover, Decide, Build, Ship, Measure, Amplify.Read the chapter

Part of

Enterprise AI Agents

About the author

Falk Gottlob

Falk Gottlob

Product Executive · Founder, Falkster.AI

Thirty years shipping product, from Microsoft Research and Adobe to Salesforce, where he grew Quip into what became Slack Canvas. Four startup exits, including a $6.5B healthcare merger and a company Microsoft bought. Four-time Chief Product Officer. Now founder of Falkster.AI, building Heidi, the company brain for AI-native businesses. This notebook is written from inside the build, not above it.

Sign in with LinkedIn to leave a comment.

Sign in with LinkedIn