LangChain's 'company harness': five pieces behind its internal agents

LangChain's post is less a product launch than a field report from running internal agents in production, and its framing is notable: it treats the hard parts as workplace plumbing rather than model capability. The company describes two common industry approaches to agent identity — acting on behalf of the invoking user, or using a fixed set of credentials — and says it ended up mixing them agent by agent rather than picking one.

The three agents illustrate that split. OpenSWE opens pull requests under the user's GitHub handle but runs background code review with its own identity. GTM Agent uses a fixed identity for Salesforce so not everyone has to log in, but only offers its on-behalf-of email tools in DMs, not shared channels. Content Studio acts for the user in Notion while posting to the company Twitter account through fixed credentials, restricted to a whitelisted set of users.

LangChain also flags memory as the least explored area, with two tiers — global and user-level — and says it has a lot of work left there. The other primitives are more settled: agents execute code outside their own sandbox by connecting into it, and a large share of runs are scheduled or event-triggered rather than kicked off by a person, whether set up by the agent's developer (OpenSWE listening for new pull requests) or by the user (a daily brief that created its own cron).