How sure an agent is
Let a Blind spots agent make the fixes it's sure of, say 90% or more, and leave the rest for people to review.
Instead of choosing between suggesting every fix and making every fix, you can tell a Blind spots agent how sure it has to be. Set it to 90%, and it makes the fixes it's at least 90% sure of itself and suggests the rest for people to review.
Set it
In Admin β Blind spots agents, Edit the agent. Under How it fixes things, choose one:
- π Suggest every fix: people approve or dismiss everything. Start here.
- π― Make the fixes it's sure of: choose how sure, from 50% to 99%. It makes the fixes at or above that itself, and suggests the rest.
- π€ Make every fix: however sure it is.
Once people have approved most of a new agent's suggestions, the agents page offers to let it make the fixes it's 90% or more sure of. The agent's badge shows its setting, such as π― 90%+.
How sure it is
Each finding waiting for review shows how sure the agent is, such as 72% sure.
- Knowa's links, tags and corrections carry Knowa's own confidence. Knowa says how sure it is when it reads an item, and drops anything it's less than 50% sure of. A link Knowa is 93% sure of (a VPN guide and the firewall it sets up) is made; a tag it's 72% sure of waits for a person.
- Links and tags found by a rule: a password whose sign-in address is a device's address is 95%, a device's tag passed on to a password for it 95%, and to a document about it 85%. A link found only because a document mentions a device's name is 60%, or 55% for a person's name.
- Gaps it can see for itself are near certain: a client with no overview (95%), an important device nobody documented (90%), a solved ticket nobody wrote up (85%), a recurring issue with no runbook (80%).
- Misconfigurations come from your tools' data, so they're 85% to 95%. The agent opens a ticket only if it has a PSA channel.
Confidence isn't a guarantee. It's the agent's best judgement, so start high and lower it as you see what it gets right. Everything it does is under Blind spots β Done, in the audit log, and documents keep their old versions.
When this is turned on, Knowa reads items again over the next while to say how sure it is of what it found before; until it has, those links and tags count as 70%.