Budget the context before adding tools
Every tool definition costs tokens on every turn. Most agent slowdowns are a context problem, not a model problem.
Playbooks
Repeatable decisions. Problem, context, approach, tradeoffs.
Idempotency before retries
Adding retries to a non-idempotent endpoint converts a visible failure into an invisible one.
Measure before caching
A cache added without a measurement is a correctness risk you took for an unknown gain.
One writer per table
The cheapest way to keep a shared database from becoming a distributed monolith.
Outbox before dual writes
If a write has to land in two places, put it in one place and let a reader fan it out.
Read replicas before sharding
When read load is the bottleneck, the cheapest correct answer is almost never sharding.