Sale Price:
$0.00
Original Price:
$8.88
An agent is not a chatbot with a new name. It acts - it looks up accounts, issues refunds, sends emails - and each choice opens more choices, until nobody can review every path. A simple billing-dispute agent has thousands. This primer is one picture to think with: the agent as a tree of paths, governed the way a bonsai is tended.
Cut, and wire. Cut the branches that must never happen - the permission that was never granted cannot be argued with, tricked, or talked past. Then hold the rest with 3 kinds of wire. Rigid wire: rules coded into the tools, not prayers written into the prompt. Supple wire: a model posted as judge - useful, and a second tree you now have to govern. And your own hands: evaluation and testing that reads the path instead of the answer, repeats every run, and probes through everything the agent reads.
Along the way: why 95% reliability per step becomes 60% over ten steps, why a rule in the prompt is a request while a rule in the tool is a control, why fluent wrongness passes a fluency check, and the sign-off questions that cannot be answered with adjectives - show me the trajectories; what happens if the model ignores this control; what is the success rate across 8 repeated runs?
About 40 pages, with figures for nearly every idea. Written by the author of Singapore's AI risk management guidelines and deliberately general: the ideas apply to agents anywhere, not just in finance.
This primer is the gateway to the fuller book - *AI Risk Management: Agentic AI*, the supervisory expectations, registers, evaluation practice and worked examples behind this way of thinking - planned for 4Q 2026.
An agent is not a chatbot with a new name. It acts - it looks up accounts, issues refunds, sends emails - and each choice opens more choices, until nobody can review every path. A simple billing-dispute agent has thousands. This primer is one picture to think with: the agent as a tree of paths, governed the way a bonsai is tended.
Cut, and wire. Cut the branches that must never happen - the permission that was never granted cannot be argued with, tricked, or talked past. Then hold the rest with 3 kinds of wire. Rigid wire: rules coded into the tools, not prayers written into the prompt. Supple wire: a model posted as judge - useful, and a second tree you now have to govern. And your own hands: evaluation and testing that reads the path instead of the answer, repeats every run, and probes through everything the agent reads.
Along the way: why 95% reliability per step becomes 60% over ten steps, why a rule in the prompt is a request while a rule in the tool is a control, why fluent wrongness passes a fluency check, and the sign-off questions that cannot be answered with adjectives - show me the trajectories; what happens if the model ignores this control; what is the success rate across 8 repeated runs?
About 40 pages, with figures for nearly every idea. Written by the author of Singapore's AI risk management guidelines and deliberately general: the ideas apply to agents anywhere, not just in finance.
This primer is the gateway to the fuller book - *AI Risk Management: Agentic AI*, the supervisory expectations, registers, evaluation practice and worked examples behind this way of thinking - planned for 4Q 2026.
Image 1 of 1