Ask HN: How are you controlling AI agents that take real actions?
We're building AI agents that take real actions — refunds, database writes, API calls.
Prompt instructions like "never do X" don't hold up. LLMs ignore them when context is long or users push hard.
Curious how others are handling this:
- Hard-coded checks before every action?
- Some middleware layer?
- Just hoping for the best?
We built a control layer for this — different methods for structured data, unstructured outputs, and guardrails (https://limits.dev). Genuinely want to learn how others approach it.
Prompt instructions like "never do X" don't hold up. LLMs ignore them when context is long or users push hard.
Curious how others are handling this: - Hard-coded checks before every action? - Some middleware layer? - Just hoping for the best?
We built a control layer for this — different methods for structured data, unstructured outputs, and guardrails (https://limits.dev). Genuinely want to learn how others approach it.