Claude CoWork is genuinely impressive - in the right context.

If you want exploration, ideation, or one-off outputs, it delivers. But the moment you move into repeatable, operational workflows, the cracks show.

I have been testing a straightforward use case:

- Reviewing a set of podcasts.
- Transcribing where necessary or using existing transcripts.
- Generating a weekly summary in a defined domain.
- Staggering processing to optimise token usage.
- Emailing the output on a schedule.

This is not complex orchestration. Yet in practice, it consistently fails to execute reliably.

Where it breaks:

- Unreliable Emailing: Drafts, failures, or silent drops.
- Authentication Issues: Token-based data access does not persist, even when credentials are re-supplied each run.
- Logic Gaps: Basic conditional logic struggles; simple "if this, then that" fails.
- No State Persistence: There is no true continuity between runs.
- Inconsistent Environments: The same task produces different behaviour across different runs.
- Scheduling Failures: Jobs trigger but do not complete.
- Opaque Errors: Failures are not surfaced clearly and often go unnoticed.
- Lack of Self-Healing: Repeated issues require manual intervention every time.

The gap here isn't intelligence. It is reliability.

Until these systems can maintain state, execute deterministically, handle credentials cleanly, and recover from failure autonomously, they remain prototypes rather than production tools.

After a month of effort (including all the frustrating "you have run out of credits while informing me how the instructions I developed failed, please wait 6 hours"), the conclusion is clear: for operational workflows, it is not yet ready for primetime.