# LangChain's Five Components for Building a Company Harness

> LangChain's internal agents across marketing, sales, and engineering share a common recipe: Slack-native UX, scoped identity, two-tier memory, sandboxed code, and scheduled runs.

- Publication: Thoai Dispatch
- Canonical URL: https://vokythoai.dev/article/langchain-five-components-company-harness
- Section: News
- Author: LangChain Blog
- Published: 2026-11-20
- Tags: langchain, agents, identity, memory, sandbox, slack, internal-tools, deep-agents

LangChain has published an account of how it built internal agents for marketing, sales, and engineering, and argues that the same five primitives recur across all of them. The post, [The five components of building a "company harness"](https://www.langchain.com/blog/components-building-company-harness), describes three agents: GTM Agent (go-to-market), OpenSWE (engineering), and Content Studio (marketing). All are built on Managed Deep Agents, which sits on top of the open source Deep Agents harness.

## Why it matters

The post frames the "company harness" as a general recipe rather than a one-off integration. The authors say the agents "completely changed how those teams work" but were "not easy to get right," and that every team landed on the same core primitives. For engineering teams evaluating internal agent platforms, the message is that the hard parts are not model quality but identity, memory, and execution boundaries.

## Technical details

The five components, as listed in the post:

1. **It lives where people already work.** For LangChain that is Slack. Each agent has its own Slack app and can be DM'd or tagged in a channel. The team added a "breakout" feature so an agent can move into its own thread rather than polluting a human conversation. Some experiences, such as reviewing a large backlog of emails, were moved to a custom UI instead.
2. **Identity.** The post calls this "the thorniest and trickiest problem." Two industry patterns are named: on-behalf-of (the agent acts as the invoking user) and fixed credentials (a service account). OpenSWE acts on behalf of the user when messaged, opening PRs with that user's GitHub handle, and with its own GitHub app identity when running in the background. GTM Agent uses a fixed identity for tools like Salesforce and on-behalf-of for email, with email tools available only in DMs. Content Studio uses on-behalf-of for Notion and fixed credentials for the company Twitter account, with a whitelist controlling who can post.
3. **Memory.** Two tiers: global memory that applies to all users, and user memory scoped to one person. Agents decide which facts go to which level. The post calls this "probably the least explored area so far."
4. **Code execution.** All three agents write and execute code. They run outside a sandbox and connect into it to execute. Each has a slightly different base image; OpenSWE selects a base image based on channel or user. All use the secret proxy on LangSmith Sandboxes to reach authenticated services safely from inside the sandbox.
5. **Scheduled runs.** A large percentage of runs are scheduled or event-triggered rather than user-invoked. Some are set by the agent developer, such as OpenSWE reviewing each new PR; others are set by users, such as a daily brief the GTM Agent created its own cron for.

## What to watch

The post positions Managed Deep Agents as providing infrastructure for channels, crons, sandboxes, memory, and identity while leaving agent behavior customizable. Teams building similar harnesses should note the identity split: on-behalf-of for user-facing actions, fixed credentials for shared services, and channel-aware restrictions such as disabling email tools outside DMs. The memory design is explicitly immature, so expect changes. The reliance on scheduled and ambient runs also implies that observability and audit of background actions matter as much as interactive ones.

## Sources

- [LangChain Blog](https://www.langchain.com/blog/components-building-company-harness) (2026-11-17)
