What does it mean that agents must be specific and testable?

Vague agents do not fire reliably — specificity is required.

Example: Compare two agents designed to improve post-meeting follow-through. Agent A: 'I will be better about following up after meetings.' Agent B: 'When a meeting ends and I was a participant, I will open my notes app within 2 minutes and write exactly three bullet points: one decision made, one action assigned to me, and one open question.' Agent A has no trigger specificity (which meetings?), no condition (what counts as a meeting worth following up on?), no measurable action (what does 'better' look like?), and no testability (how would you know if it fired?). Agent B specifies the trigger (meeting ends), the condition (I was a participant), the action (three categorized bullet points in a specific app), and the time constraint (within 2 minutes). After one week, you can count: how many meetings did I attend, and how many times did the three-bullet-point protocol execute? If the ratio is below 80%, the agent is failing and you can diagnose why. Agent A cannot be diagnosed because it cannot be measured.

Try this: Select one agent you currently run — a rule, habit, or protocol you follow in recurring situations. Write it down exactly as it exists in your mind right now. Then apply the specificity test: (1) Can you identify the exact trigger — the observable event that should activate this agent? (2) Can you state the condition — the criteria that must be true for the agent to fire? (3) Can you describe the action in enough detail that someone else could execute it identically? (4) Can you define what success looks like in measurable terms? For each dimension that fails the test, rewrite it with concrete specificity. Then define your falsification criterion: what observable outcome over the next seven days would prove this agent is not working? Run the agent for one week and check the data.