The irreducible epistemic atoms underlying the curriculum. 4,828 atoms across 8 types and 2 molecules
Define agent triggers as observable external events or measurable internal states rather than subjective feelings or abstract conditions, because vague triggers cannot be recognized reliably when they occur.
Set agent review cadences at 7 days for new habits, 30 days for established behaviors, and immediately after any major context change, because review timing must match the actual rate of drift in each agent type.
Define agent success as a measurable outcome with a minimum acceptable firing rate threshold (typically 80% over one week for new agents) rather than subjective satisfaction, because subjective assessment systematically inflates reliability perception.
Externalize critical reliability agents (medication, safety checks, high-stakes commitments) to tools or environments rather than trusting biological memory, because internal agents degrade precisely when stakes are highest—under stress and cognitive load.
Use hourly momentary sampling over 48+ hours rather than end-of-day recall when auditing behavioral agents, because retrospective memory systematically overweights salient successes and underweights invisible failures.
Classify each observed behavior as designed (you can identify the installation decision) or default (no identifiable decision point) during audits, treating the classification question itself as a detection mechanism for unexamined automation.
When an agent fires below 80% of expected opportunities over 30 days, reduce it to the simplest executable version before adding any complexity, because unreliable agents cannot be improved through sophistication.
Design minimal viable agents to execute in under two minutes with zero preparation before attempting multi-step sequences, because automaticity requires low activation energy and activation energy must be minimized before sophistication is added.
When agent sub-behaviors can execute independently without logical dependency, separate them into distinct agents with independent triggers rather than coupling them into sequences, because coupled agents produce cascading failures.
Document every agent in a structured five-component format: (1) Name, (2) Trigger, (3) Conditions, (4) Actions, (5) Success criteria, to enable systematic review and prevent silent degradation.
Write agent action steps as specific ordered procedures rather than aspirations or principles, requiring sufficient granularity that someone unfamiliar could execute them without clarification.
Maintain a failure log where every agent misfire is recorded with date, agent name, what happened, and hypothesis about why, then review weekly to extract patterns.
When an agent fails, diagnose which component broke—trigger (never activated), condition (activated but context wasn't right), or action (executed but too vague/complex)—before attempting any redesign.
Fix only one component (trigger, condition, or action) per agent iteration rather than redesigning multiple components simultaneously, to maintain causal attribution of what changes produced which effects.
Before building any agent, explicitly name the schema it operates on by writing what the agent assumes about how the world works, because unexamined schemas produce systematically wrong outputs despite reliable execution.
After any difficult social interaction, complete a 5-minute written debrief comparing what your social agent prescribed versus what you actually did, documenting the gap to update the agent based on real behavioral data.
When a decision situation represents a once-in-career strategic pivot rather than a routine recurrence, suspend your decision agent and return to full deliberation even if the agent would produce an answer.
After each decision agent activation, update the agent's criteria based on whether they produced a good outcome, treating the agent as a living heuristic that improves with each use rather than a permanent law.
For sleep agents, set a hard time-based trigger (like 9:30 PM) rather than a subjective state trigger (like 'when I feel tired'), because fatigued brains systematically misjudge their own tiredness.
Build nutrition agents at the meal preparation layer (Sunday evening meal prep) rather than the consumption layer (what to eat when hungry), because deciding what to eat while hungry and facing an open refrigerator is too late to override depletion.
For stress management agents, use physiological signals (chest tightening, jaw clenching, shallow breathing) as triggers rather than cognitive assessments ('when I feel stressed'), because stress impairs the self-assessment capacity needed to recognize stress.
When discovering that your designed agents conflict with each other, resolve the conflict through documented priority hierarchies rather than case-by-case deliberation, making the resolution rule itself part of your agent system.
Build feedback loops into agent systems through regular review asking whether agents fired, whether they produced intended outcomes, and whether conditions have changed, treating review as essential maintenance not optional improvement.
Design triggers using the camera test: if a video camera could not detect the exact moment the trigger fires, the trigger is too vague to fire reliably and must be replaced with an observable event.