Notes on scaling browser agents and distributed execution systems.
Scaling agents is not about retries. It is about explicit lifecycle transitions.
Example model:
creating → active → rotating → expired → error
If lifecycle is machine-readable, coordination becomes deterministic.
Browser sessions accumulate entropy:
- stale cookies
- hidden 429s
- resource exhaustion
- inconsistent DOM state
Retries amplify entropy. Explicit expiration reduces it.
Scaling improves when executors are:
- short-lived
- observable
- restartable by design
Death should be a feature, not an unexpected failure.
Recoverable rate limits: retry + jitter.
Structural decay: expire session and recreate.
Treat them differently.
Concurrency is not linear.
Hidden state surface area grows with parallelism.
Explicit state transitions reduce cascading failure.
- Should browsers expose lifecycle hooks natively?
- How to measure session entropy?
- What is the optimal rotation boundary?
This repository is a collection of evolving notes.