Checking the agent with something that isn't the agent sounds obvious. I'd guess most small teams don't have that, though. They read the agent's own summary and move on.
Steve, the distinction between authenticating an agent and verifying its intended action is particularly important.
A customer's permission to access an account should not automatically authorise every action the agent might infer from a broad instruction.
I would want the authorisation record to specify the task, permitted actions, limits and expiry, with an independent system checking execution against that scope.
The difficult question is how to support legitimate changes in a multi-step workflow without constantly interrupting the user.
How would you design that balance between verifiable intent and practical usability?
Signal 09 and the 100-action test in Signal 03 describe the same gap from two ends.
The role of non-human identity is the interesting part. A system user acting on behalf of a person, inheriting their credentials, executes a set of instructions the same way every time. An AI agent can interpret the task differently. Humans can too, which is why a human acting on someone’s behalf works within a limited scope, takes only the actions they are authorised to take, and is usually held liable for going off-script. A fiduciary moving a client’s money, or an attorney taking legal action for a client, is the example. An AI agent, unlike a professional or a company, cannot be held liable, and the law hasn’t quite caught up on how to govern these actions yet.
In practice that pushes the question back to whoever deployed the agent, and it only works if a person is named. Suppose a support agent approves refunds under a set amount. Until now a team lead handled the exceptions. That lead decided what counted as an exception, took the call when a customer escalated, and held the refund budget. Remove the lead and the agent inherits the task. Nobody inherits those three jobs.
So when you run the 100-action test, add four columns to the record: the named owner at the time of the action, who can shut the agent off, who accepted the residual risk, and which budget carries the cost. If any of the four is blank, the log shows activity, not accountability.
For the next Boardroom Prompt: for your three most consequential agents, who can switch each one off today, and is that the same person who accepted the risk?
Checking the agent with something that isn't the agent sounds obvious. I'd guess most small teams don't have that, though. They read the agent's own summary and move on.
Steve, the distinction between authenticating an agent and verifying its intended action is particularly important.
A customer's permission to access an account should not automatically authorise every action the agent might infer from a broad instruction.
I would want the authorisation record to specify the task, permitted actions, limits and expiry, with an independent system checking execution against that scope.
The difficult question is how to support legitimate changes in a multi-step workflow without constantly interrupting the user.
How would you design that balance between verifiable intent and practical usability?
Signal 09 and the 100-action test in Signal 03 describe the same gap from two ends.
The role of non-human identity is the interesting part. A system user acting on behalf of a person, inheriting their credentials, executes a set of instructions the same way every time. An AI agent can interpret the task differently. Humans can too, which is why a human acting on someone’s behalf works within a limited scope, takes only the actions they are authorised to take, and is usually held liable for going off-script. A fiduciary moving a client’s money, or an attorney taking legal action for a client, is the example. An AI agent, unlike a professional or a company, cannot be held liable, and the law hasn’t quite caught up on how to govern these actions yet.
In practice that pushes the question back to whoever deployed the agent, and it only works if a person is named. Suppose a support agent approves refunds under a set amount. Until now a team lead handled the exceptions. That lead decided what counted as an exception, took the call when a customer escalated, and held the refund budget. Remove the lead and the agent inherits the task. Nobody inherits those three jobs.
So when you run the 100-action test, add four columns to the record: the named owner at the time of the action, who can shut the agent off, who accepted the residual risk, and which budget carries the cost. If any of the four is blank, the log shows activity, not accountability.
For the next Boardroom Prompt: for your three most consequential agents, who can switch each one off today, and is that the same person who accepted the risk?