Notes from production
A knowledge base that audits itself
Every morning a set of scripts reads the whole knowledge base and the last day of agent output, tests each agent against its own rules, and posts five lines to the team channel.
Agents never hold your API keys
When an agent needs data from your CRM, accounting or inbox, we put a small service between them that holds the credential and hands over only the fields you approved.
Drafts only by default
Every agent we ship does the grind on its own, and none of them can send, post, pay or publish until a person says so.
Rehearse a year of agent behaviour in 22 minutes
Before a change reaches a production fleet we run it through an isolated copy of the live setup, against the real models, and read what the agents say.
Spend tokens only on judgement
If a script can decide a step with certainty the model should never see it, and for the steps that are left the model should see the smallest slice of evidence that settles them.