Build vs Call Another Agent
Build vs Call Another Agent
You need your agent to send emails. Do you build email sending into your agent, or call an external email agent through an API? This decision comes up for every new capability and most people get it wrong.
The 3x/Day Rule
If you need a capability three or more times per day, build it in. The integration overhead of calling an external agent - authentication, error handling, format translation, retry logic - accumulates with every call. At three calls per day, you spend more time maintaining the integration than it would cost to build the feature natively.
External agent calls look cheap on paper. One API call. But then you need to handle timeouts. The external agent's response format changes. Their rate limits hit during your busiest hour. Their service goes down on Friday afternoon. Each edge case adds code, adds complexity, adds fragility.
Integration Overhead Is Always Higher
People estimate integration at "a few hours." It is never a few hours. You need to handle authentication and token refresh. You need to normalize the response format into your agent's internal representation. You need error handling for every failure mode. You need monitoring to know when the integration breaks silently.
For a capability you use once a week, this overhead is acceptable. For something your agent does constantly, it is a tax on every operation.
When External Agents Make Sense
External agents make sense for specialized, infrequent tasks. Running a complex data analysis once a month. Generating a specific report quarterly. Tasks where the external agent has deep domain expertise you would spend months replicating.
The decision framework is simple: frequency times complexity. High frequency and low complexity - build it. Low frequency and high complexity - call it. Everything else requires judgment, but err toward building.
Fazm is an open source macOS AI agent. Open source on GitHub.