Thesis
Agents fail on unanswerable questions
What did this cost. What is it allowed to touch. Who approved the destructive step. What ran last Tuesday. Agents get into production trouble when nobody can answer these — the gap between an agent acting and a human knowing is discovery lag, and it is the thing to engineer down.
Deixic answers the operating questions with spend by agent, a live agent list, approve/deny on gated work, connected tools, and an activity trail.
Five things that keep breaking
- Agents reach across data they were never scoped to. Security finds out in the postmortem.
- Destructive operations run without a real human approval — drop tables, force pushes, deletes.
- There is no single place to stop a running agent and its in-flight credentials.
- The SIEM has the API call but not the operator and run that caused it.
- Nobody can query “what can this agent reach?” before go-live.
The five in depth: the field-notes version. What the app does today: What you can do. The rest of the doctrine: the blog.