Concepts
The irreducible epistemic atoms underlying the curriculum. 4,828 atoms across 8 types
Measure error rates over time as system health indicators…
Measure error rates over time as system health indicators rather than treating individual errors as isolated moral failures requiring emotional response.
Deliberately introduce controlled stressors and errors…
Deliberately introduce controlled stressors and errors during training to build redundant pathways and prevent overfitting to ideal conditions.
Design shared information substrates between agents…
Design shared information substrates between agents by specifying what each agent reads from and writes to, rather than granting unrestricted access to all information.
When two cognitive agents both claim authority over the same
When two cognitive agents both claim authority over the same resource or decision, treat the conflict as a structural architecture problem requiring scope redefinition rather than a personal discipline failure.
Define context-specific priority orderings rather…
Define context-specific priority orderings rather than global rankings, because the relative importance of competing values and goals shifts with context (work hours vs. personal time vs. crisis mode).
Before retiring an agent, map all upstream and downstream…
Before retiring an agent, map all upstream and downstream dependencies by identifying which systems consume its outputs or assume its existence, then specify whether each dependency will be transferred, consciously dropped, or temporarily covered.
Prevent cognitive thrashing by defining agent execution…
Prevent cognitive thrashing by defining agent execution sequences before starting work rather than discovering the correct order through repeated failed attempts and task-switching.
Distinguish between genuinely dependent tasks that must…
Distinguish between genuinely dependent tasks that must execute sequentially and independent tasks that only appear sequential due to implicit habit or convenience, running independent tasks in parallel to compress total completion time.
Allocate optimization effort to reducing serial bottlenecks…
Allocate optimization effort to reducing serial bottlenecks rather than accelerating already-parallel work, because the serial fraction determines the maximum possible speedup regardless of parallelization efforts (Amdahl's Law applied to cognition).
Serialize novel, complex, or unfamiliar cognitive tasks…
Serialize novel, complex, or unfamiliar cognitive tasks that require deliberate System 2 processing, while allowing well-practiced, habituated tasks to run in parallel through automatic System 1 processing.
Align the mental models underlying different agents'…
Align the mental models underlying different agents' operations to enable implicit coordination, because agents operating on different assumptions about goals, constraints, and task structure will produce locally rational but globally incoherent outputs.
Include minimal sufficient context with every inter-agent…
Include minimal sufficient context with every inter-agent message—what the previous agent did, what constraints apply, and what the receiving agent needs to know—rather than transmitting either everything or just the artifact.
Specify completion criteria when one agent hands work…
Specify completion criteria when one agent hands work to another, defining what 'done' looks like for the receiving agent, to prevent both over-delivery and under-delivery failures.
Activate the orchestrator agent only at transition points…
Activate the orchestrator agent only at transition points and sequence boundaries, not during execution within individual agents, to avoid converting coordination into rumination.
Build orchestrator protocols that survey system state…
Build orchestrator protocols that survey system state (energy, constraints, deadlines, priority), apply sequencing heuristics, and then release control to the selected agent without continuing supervision.
Externalize the output of one cognitive agent…
Externalize the output of one cognitive agent into structured, persistent format before transitioning to the next agent, because working memory context decays and does not transfer automatically between cognitive modes.
Enumerate all cognitive agents in your system, identify what
Enumerate all cognitive agents in your system, identify what each consumes and produces, and draw directed edges from producers to consumers to make dependency structure explicit and debuggable.
Protect agents with high fan-out (many downstream…
Protect agents with high fan-out (many downstream dependents) as force multipliers, and treat agents with high fan-in (many upstream dependencies) as fragility points requiring redundancy.
Distinguish structural dependencies (Agent B cannot run…
Distinguish structural dependencies (Agent B cannot run without Agent A's output) from accidental dependencies (Agent B follows Agent A by habit) and eliminate accidental dependencies to increase parallelism.
Define allocation rules for contested resources before…
Define allocation rules for contested resources before contention occurs rather than negotiating access during execution.
Allocate larger time blocks to fewer concurrent tasks rather
Allocate larger time blocks to fewer concurrent tasks rather than fragmenting time into many small slices, because context-switching costs consume 20-40% of effective capacity.
Select collaboration patterns by analyzing task dependencies
Select collaboration patterns by analyzing task dependencies first—pipeline for sequential dependencies, fan-out for independent tasks, consensus for reciprocal dependencies—rather than by preference or habit.
Build transactive memory within teams—shared knowledge…
Build transactive memory within teams—shared knowledge of who knows what—to reduce coordination overhead by enabling efficient delegation rather than requiring consensus on every decision.
Never add agents to sequential reasoning tasks—multi-agent…
Never add agents to sequential reasoning tasks—multi-agent coordination on sequential dependencies increases overhead without increasing throughput.