Workflows are Agents that Work

I've got a friend who is very into crystals and herbs for healing. Once, another friend muttered to him: "industrial medicine is herbal healing - it's just what we call the herbs that work."

I think there's something there that applies to workflows and agents too.

Agents certainly represent a holy grail that just "figures things out," but post hoc, any one process an Agent figures out is essentially a static workflow. Just one the Agent authored instead of you.

Like most things in life, agent requests will have a Zipfian to them. 80% of the volume will be agglomerated within 20% of the distinct requests.

And so even for agent platforms, it makes a lot of sense to think about how to officiate and perfect the most popular workflows.

A workflow perspective is inevitable

For the vast majority of agent-based startups, here's what the B2B deal looks like:

  • Startup: Our agent can do anything!
  • Customer: Cool, but we want to use it to do this one thing.
  • Startup: (Does a 1-month pilot..) Our agent is getting an 80% success rate on your use case!
  • Customer: I'll sign if you make that 95%
  • Startup: Adds extra-LLM features, UI, notifications, observability, and error handling specific to the customer's use case

Until AGI, that's how most agent startups will survive. Post AGI, they're probably cooked anyway, since their need will be obviated by the foundation model companies themselves.

© 2025 Edward Benson. All rights reserved.