Small Expert Cohort Produces Most Output

Claude Cowork is still behaving like a prosumer power tool, not a mainstream team operating system. At Scale AI, the heaviest users are project managers, product managers, and developers who can debug broken tool chains, manage permissions, and keep multi step workflows alive when they fail. That matches a broader pattern where adoption spreads first through a small group willing to wire connectors, test prompts, and babysit automations until they become reliable.

• At Scale AI, the bottleneck is not raw model quality, it is operational complexity. Single tool workflows and narrow QC automations can get reliable enough to use heavily, but once four or five tools need to pass context across Slack, Airtable, Monday, Linear, and internal systems, errors cascade and only technical users tend to recover them.
  1 sacra
• A comparable pattern shows up in another heavy user setup, where one operator runs about twenty daily Codex automations, but still spends about 14% of usage maintaining integrations, auth, and custom glue code. That is what early agent adoption looks like in practice, lots of value, but concentrated in users who tolerate software maintenance.
  3 sacra
• The contrast case is narrower ops automation. At Whop, Cowork can run daily Slack and Gmail based reporting with near full autonomy after a week of testing, but the operator still keeps humans on final external emails and sensitive compliance or money workflows. Adoption broadens fastest where tasks are repetitive, internal, and easy to verify.
  4 sacra

The next phase is a shift from agent capability to product design. As these tools hide GitHub, auth, and debugging behind simpler approvals, templates, and audit trails, the output curve should flatten, with less volume concentrated in a tiny expert class and more of it becoming normal team workflow.