I’ve been experimenting with agent-style architectures in environments where
cloud assumptions don’t hold: single-digit MB memory, near-zero startup time,
no background services.
What surprised me is not what’s hard, but what breaks immediately:
– dynamic planning
– tool abstraction layers
– prompt-heavy orchestration
– anything assuming async I/O is “cheap”
In practice, the architecture ends up looking very different from most
agent frameworks discussed today.
For people who’ve worked on embedded systems, edge inference, or constrained
runtime environments: what design assumptions did you have to unlearn first?
For anyone asking for concrete context, I’ve been experimenting with a small
reference implementation that explores these constraints end-to-end.
Linking here as supporting material, not a tutorial:
https://github.com/nullclaw/nullclaw
What surprised me is not what’s hard, but what breaks immediately: – dynamic planning – tool abstraction layers – prompt-heavy orchestration – anything assuming async I/O is “cheap”
In practice, the architecture ends up looking very different from most agent frameworks discussed today.
For people who’ve worked on embedded systems, edge inference, or constrained runtime environments: what design assumptions did you have to unlearn first?