Here is a question worth sitting with.
If something unusual happened in your last AI session, would you know? Not in theory. Right now, with your current setup, do you have a record of which agents ran, what they accessed, and what they produced?
For most practitioners, the honest answer is no.
Anomaly detection is not a sophisticated monitoring system. It is knowing what your system normally does well enough to notice when it does something different. That knowledge only comes from a log. And the log has to exist before you need it, not after something goes wrong.
The instinct most people have is to create a log file and update it manually after each session. That instinct is wrong. Manual logging fails because it depends on your memory and discipline at the end of a long working day. An inconsistent baseline is worse than no baseline. It creates false confidence.
The right approach is to build logging into the commands your AI system already runs.
Your session-opening command runs every session. Add a logging step that records the scope and authorized clusters before any work begins.
Your commit command executes your highest-risk action. Add a pre-commit log entry that records exactly what is being committed and confirms the human authorization before it executes.
Your session-closing command ends each session. Add a logging step that writes every significant action taken, timestamped, to your action log automatically.
You authorize the work. The AI records it. No extra steps.
After 30 days of consistent logging, four anomaly signals become immediately visible from the data:
Scope creep: an agent requests access outside its designated cluster.
Repetition spikes: the same action attempted more than three times in a single session.
Instruction source anomalies: an agent receiving instructions from somewhere other than the orchestrator or you directly.
Confirmation bypass attempts: a commit or external action attempted without your authorization.
There is a larger point here for those deploying for clients.
The EU AI Act requires automated audit trails for AI systems classified as high risk. Canadian regulators are watching that framework closely. Your action log, built properly from day one, is the beginning of the compliance artifact that will eventually be required of every serious AI practitioner.
Start it now. Build it into your existing commands. By the time it is required, you will have months of clean data behind you instead of scrambling to reconstruct what your agents did.
If you reviewed your last 30 days of AI sessions right now, could you account for every significant action your agents took?
Tom Olakitan is the founder of Cerebrium Consulting. He helps BC-based SMBs build AI operating systems that are secure enough to trust and practical enough to use.