Series 001 · practical agent safety
Do not give an agent a dangerous tool and call the prompt a fuse.
Agents are useful when they can touch the work. They are dangerous for exactly the same reason. This series is the public-safe version of the discipline: boundaries first, autonomy second.
The System Prompt Is Not the Safety System
Why behavioural text is not the same thing as a boundary.
Why Useful Agents Need Verifiers
A worker without an independent check is just a confident risk amplifier.
The Real Problem With AI Agents Is Context Selection
Most failures start before the model answers: wrong evidence, wrong tools, wrong memory.
Why Prompting Is Becoming Spec Writing
A proper prompt is closer to a work order than a magic phrase.
What I mean by safe enough to use
- No live customer, payment, deployment or purchasing action without an approval gate.
- Every tool permission should have a reason, a log and a smaller alternative.
- The reviewer must be outside the worker path when the action can hurt money, reputation or safety.
- Rollback is part of the design, not something invented after damage.