We are handing authority to systems that act faster than we can watch, on credentials we hand them, and that forget their own reasoning the moment a task ends. When an agent does something that matters, moves money, changes a record, acts on behalf of an agent acting on behalf of a person, the questions show up after the fact: what happened, who is accountable, and on whose authority. Most agent stacks answer with application logs. Those logs are mutable, reconstructed later, and only as trustworthy as whoever owns the server. That is not good enough.

This talk argues that durable, tamper-evident logging is not a governance checkbox you bolt on at the end. It is a precondition for shipping agents you can stand behind. I separate logging from notary action: sign and hash-chain every consequential step the instant it happens, bind it to the responsible agent or person, and chain delegated authority all the way up so every entry resolves to its origin. A record built this way verifies offline for anyone holding the public keys. You do not have to trust the operator, the vendor, or the model provider.

I walk through the moving parts in plain terms: append-only chains, signing keys, and the choice to gate high-consequence actions while simply recording the routine ones. Then the part people miss. When the controlled path takes away the agent's guesswork about what each step requires, safety and governance stop fighting capability and become the fast path. Regulatory expectations like the EU AI Act's logging requirements fall out as a byproduct of building it right, not as a separate compliance project.

You will leave with a framework for deciding what to record, what to gate, and how to produce an account of agent behavior that holds up when someone finally asks what happened.