LangChain runs its sales, engineering and marketing agents on the same five parts. Identity was the hardest one

Whose name does your agent sign with when it sends the email? Decide that before you build anything else.

What happened

LangChain described how it built its internal agents: a GTM Agent that watches email, responds to new qualified leads, sends daily briefs and runs outbound, OpenSWE for drafting and reviewing pull requests, and Content Studio for marketing assets. The company says every team landed on the same five components:

1. The agent lives where people already work, which for LangChain is Slack.
2. It knows who is asking, and acts on behalf of that person for some services and with its own identity for others.
3. It has memory at two levels, global and per user.
4. Code runs in a sandbox that can reach services needing auth.
5. A large share of runs are scheduled or triggered by events, not started by a person.

LangChain calls identity the thorniest problem. In its setup, PRs opened from a chat use the requester's GitHub handle, while background code review uses the agent's own GitHub app. Salesforce uses one fixed identity so not everyone has to log in. Email acts on behalf of the user and is only available in direct messages, never in shared channels. The company Twitter account uses fixed credentials, and only a whitelisted set of users can post. They also built a custom UI for reviewing large backlogs of draft emails, because doing that in Slack was not a good experience.

My take

This is the checklist I wish every client asked for before their first agent. The model is the easy part. The questions that cause trouble later are:

1. Who is the agent acting as? Write it down per tool: the requester, a service account, or a shared company login.
2. Where can each tool be used? Anything that sends email or posts publicly should be off in shared channels.
3. Who can trigger the risky actions? Reading and drafting can be open. Posting and sending should be a short list of people.
4. What runs on a schedule? Ambient runs like daily briefs need logs someone actually reads.

Small teams can do this in n8n or Zapier with separate credentials per workflow and a Slack approval step. It is a few extra hours at the start and it saves a very awkward conversation later.