This maps cleanly to clinical agents because a bare loop makes the action boundary inspectable.
In an OR-adjacent workflow, I would want the first version to stay just as small: read a known source, detect one mismatch, route it to a role, and stop before it claims readiness.
The loop is useful only if each tool call shows what it observed, what it was allowed to touch, and what stayed forbidden.
The harness choice mirrors a cost-structure decision I see in financial planning: serverless (Modal) vs. token-based APIs is like variable vs. fixed costs.
For a side project, variable cost keeps burn low until scale. For a startup with predictable load, fixed-cost APIs cap the downside.
Cost-aware design matters as much as code quality.
This maps cleanly to clinical agents because a bare loop makes the action boundary inspectable.
In an OR-adjacent workflow, I would want the first version to stay just as small: read a known source, detect one mismatch, route it to a role, and stop before it claims readiness.
The loop is useful only if each tool call shows what it observed, what it was allowed to touch, and what stayed forbidden.
Exactly!
Bare-bones is the right place to start. The loop stays this small forever, everything after it is tools, gating, and context.
I built awesome-agent-architecture on that idea, each section adds one mechanism to the same runnable loop.
https://github.com/hardness1020/awesome-agent-architecture
Nice resource
The harness choice mirrors a cost-structure decision I see in financial planning: serverless (Modal) vs. token-based APIs is like variable vs. fixed costs.
For a side project, variable cost keeps burn low until scale. For a startup with predictable load, fixed-cost APIs cap the downside.
Cost-aware design matters as much as code quality.
Yes, it does! That's why I did my best to present a mindmap on how to decide what's best for you