Log every agent misfire with date, name, event, and hypothesis — weekly pattern review turns failures into redesign data

Why This Is a Rule

Individual agent failures are annoying but uninformative. Patterns across failures are diagnostic gold. A single failure tells you the agent didn't fire today; a pattern across 10 failures tells you the agent fails specifically on days with morning meetings, or when you slept less than 7 hours, or when the trigger occurs during a context switch. The failure log converts individual data points into a pattern-detectable dataset.

The four fields — date, agent name, what happened, hypothesis about why — capture the minimum information needed for pattern extraction. The date enables temporal patterns (day of week, time of month). The agent name enables per-agent analysis. "What happened" captures the failure mode (trigger missed? condition blocked? action too hard?). The hypothesis about why captures your in-the-moment theory, which is often more accurate than retrospective reconstruction because you're closer to the actual context of failure.

Weekly review is the extraction cadence: frequent enough to catch patterns before they become entrenched, infrequent enough to accumulate sufficient data points for pattern visibility. A single failure per week might be noise; five failures of the same agent on the same type of day is a clear signal.

When This Fires

• Every time a behavioral agent fails to fire when it should have
• During weekly reviews when extracting patterns from the accumulated log
• When an agent seems unreliable but you can't articulate why — the log will show you
• Before diagnosing a failing agent (Diagnose before redesigning — identify whether trigger, condition, or action broke before changing anything) — check the failure log for patterns first

Common Failure Mode

Not logging failures because they feel like personal shortcomings: "I don't want to write down that I failed again." The log isn't a shame record — it's a diagnostic tool. Software engineers don't feel ashamed of error logs; they use them to fix bugs. Agent misfires are bugs in your behavioral system design, not character failures.

The Protocol

(1) Keep a simple failure log — notebook, app, or spreadsheet. Four columns: Date, Agent Name, What Happened, Why (hypothesis). (2) When an agent misfires, log it as close to the event as possible — don't wait until evening. The hypothesis is most accurate when you're still in the context. (3) Weekly: review the log. Look for patterns: which agents fail most? On which days? Under what conditions? What hypotheses recur? (4) When a pattern emerges, use it to redesign: if the agent fails every Monday morning, the trigger needs redesigning for the Monday context. If it fails whenever you're tired, the action needs simplification (When an agent fires below 80% after 30 days, simplify before sophisticating — unreliable agents need reduction, not enhancement). (5) The failure log is the raw data; the weekly review is the analysis. Neither works without the other.