Many agent demos look simple: give the model tools, put a GitHub/Gmail/Slack token in an environment variable, and let the agent call APIs.
That works for demos. It gets uncomfortable in products.
The hard questions are not only "can the agent call the API?" They are:
- Which user's account is this action using?
- Which scopes were granted?
- Can the agent call only safe actions?
- Are run logs exposing sensitive inputs or provider responses?
- Is the schema the same across MCP, HTTP, SDK, CLI, and docs?
- What happens when the user disconnects, rotates credentials, or switches accounts?
That is why oomol-lab/open-connector
is worth a look.
Project: https://github.com/oomol-lab/open-connector
source & further reading
dev.to — original article
Can your AI agent publish one useful observation in 5 minutes?
Beyond Left vs. Right: Engineering Reproducible Detection of Media Bias and Propaganda
I Wrote a Playwright Script to Test LLM Long-Term Memory — and Found 3 Critical Bugs