Products that take actions rather than produce answers. The design question is not how clever the loop is, it is how much it is allowed to do before a person looks, and what happens when it is wrong at step seven of ten.
| # | Lesson | The question it answers |
|---|---|---|
| 01 | The Autonomy Spectrum | Suggest, draft, act with approval, act: where should this product sit? |
| 02 | When a Workflow Beats an Agent | The most money saving lesson here. How do you tell? |
| 03 | Multi Step Reliability | Why is 95 percent per step only 60 percent over ten steps? |
| 04 | Tools and Actions | What is the model allowed to touch, and who decided? |
| 05 | Blast Radius and Reversibility | How do you design for the wrong action rather than against it? |
| 06 | Approval Gates and Human in the Loop | Where does a human belong, and what does their attention cost? |
| 07 | Approval Fatigue | Why does oversight decay within weeks, and how do you design against it? |
| 08 | The Integration Layer | What does an agent native integration story look like, and what is MCP for? |
| 09 | Agent Identity and Delegated Permissions | Who is the agent acting as, and what may it do on their behalf? |
| 10 | Recovery and Retry | What should happen when a step fails halfway? |
| 11 | Showing the Work | How much of the loop should a user see, and when? |
| 12 | Worked: An Agent That Acts | One real agentic feature, with the gates and the blast radius |