What AI Agents Are Actually Worth Building?
What AI Agents Are Actually Worth Building?
The AI agent gold rush has teams building general-purpose assistants that do everything poorly. The agents actually worth building target specific repetitive workflows and do them well.
The General Assistant Trap
"An AI agent that can do anything" sounds compelling until you realize it means an agent that does nothing reliably. General-purpose agents spread their capabilities thin, handle edge cases poorly, and require extensive prompting to do basic tasks.
The agents that users actually keep using are narrow and opinionated:
- An agent that processes your expense receipts every week (not "a financial assistant")
- An agent that reformats data between your CRM and your spreadsheet (not "a data integration platform")
- An agent that drafts follow-up emails after meetings using your notes (not "a communication assistant")
How to Identify Worth-Building Agents
A workflow is worth automating with an agent if it has:
- Repetition - you do it at least weekly
- Pattern - the steps are similar each time but not identical (if identical, a script is enough)
- Judgment - it requires small decisions that a rule engine cannot handle
- Low stakes - failure is annoying but not catastrophic
- Clear inputs and outputs - you can define what goes in and what should come out
If all five are true, build the agent. If fewer than three are true, build a script or do not automate at all.
The Value Calculation
An agent that saves you 20 minutes per day on a specific task is worth more than one that theoretically saves hours but requires 10 minutes of setup and verification each time. Reliability beats capability.
The boring agents - the ones that handle expense reports, format documents, update CRMs, organize files - are the ones that actually get used daily. Build boring agents.
Fazm is an open source macOS AI agent. Open source on GitHub.