The five components of building a 'company harness'
Chase and Broekhuizen describe five primitives LangChain found recurring across its internal department agents: embedding agents in existing workflows (mainly Slack), identity management (acting as the user vs. fixed credentials), dual-tier global/per-user memory, secure sandboxed code execution with secret proxies, and scheduled or event-triggered runs. They note a significant share of agent runs happen ambiently via cron or triggers rather than direct user invocation.
What this could change
Editorial: a concrete, field-tested checklist for anyone deploying internal agents across a company's teams.
Perspective & interests
LangChain sells the agent infrastructure this framework implies, so codifying the 'company harness' pattern promotes its products.