Principles
2,201 principles · page 46 of 46
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.
Make coordination activities visible and calculate the ratio
Make coordination activities visible and calculate the ratio of coordination time to productive time, setting an explicit budget that new coordination mechanisms must fit within.
Prefer implicit coordination mechanisms—shared conventions…
Prefer implicit coordination mechanisms—shared conventions, templates, routines, and mental models—over explicit communication, because implicit coordination consumes dramatically less cognitive bandwidth.
Reject new coordination mechanisms that exceed…
Reject new coordination mechanisms that exceed the coordination budget unless an existing mechanism is removed, treating coordination capacity as a zero-sum resource.
Protect beneficial emergent patterns by maintaining…
Protect beneficial emergent patterns by maintaining the interaction conditions that produce them rather than formalizing the emergent behavior into explicit rules that eliminate the generative interaction.
When diagnosing harmful emergent patterns, change one agent…
When diagnosing harmful emergent patterns, change one agent or one interaction context rather than adding new agents, because increased interaction density can produce unpredictable new emergence.
Before adding a new agent to an existing system, count…
Before adding a new agent to an existing system, count the interaction channels it will create — if n(n-1)/2 exceeds your coordination capacity, either remove an existing agent first or do not add the new one.
Before removing an agent, audit what consumes its outputs…
Before removing an agent, audit what consumes its outputs, what constraints it enforces, and what failures it masks — then explicitly reassign, accept, or reroute each dependency before removal.
Reduce an agent's scope or frequency before eliminating…
Reduce an agent's scope or frequency before eliminating it entirely, using the graduated period to observe which hidden dependencies reveal themselves through degraded performance.
Verify handoff integrity with evidence rather than memory…
Verify handoff integrity with evidence rather than memory — check whether upstream outputs actually arrived at downstream consumers, not whether you believe they should have.
Practice coordination sequences consciously and repeatedly…
Practice coordination sequences consciously and repeatedly until they compile into System 1 automatic execution — compilation requires repetition of the full coordination pattern, not just individual components.
Attribute apparently effortless performance to well-compiled
Attribute apparently effortless performance to well-compiled multi-agent coordination rather than innate talent — the attribution determines whether you view the capability as reproducible or fixed.
Allocate attention only to tasks where your unique judgment…
Allocate attention only to tasks where your unique judgment, expertise, or authority creates irreplaceable value — systematically offload everything else through delegation to preserve capacity for highest-leverage work.
Identify your system's constraint and optimize only…
Identify your system's constraint and optimize only that constraint — improvements to non-bottleneck activities produce no increase in throughput.
Delegate to systems (checklists, automation, tools…
Delegate to systems (checklists, automation, tools, environments) before delegating to people — systems scale without increasing cognitive overhead, people do not.
Score tasks on three dimensions — irreversibility (cost…
Score tasks on three dimensions — irreversibility (cost of error), identity-centrality (builds core capability), and cognitive uniqueness (requires your specific context) — then delegate only tasks scoring low across all three.
Never delegate identity-defining decisions (those expressing
Never delegate identity-defining decisions (those expressing core values), context-dependent judgments (requiring non-transferable accumulated knowledge), or irreversible-consequence decisions (where errors cannot be undone) — these three categories require that decision authority and accountability remain unified.
Specify every delegation with five components
concrete outcome (what 'done' looks like), constraints (boundaries that must not be crossed), success criteria (how results will be evaluated), available resources (what the delegate can use), and timeline with checkpoints (when and how progress will be verified).
Specify the desired outcome with precision while leaving…
Specify the desired outcome with precision while leaving the method to the delegate's discretion to preserve autonomy and enable adaptive problem-solving.
Communicate both the objective and the intent behind it so…
Communicate both the objective and the intent behind it so delegates can adapt when conditions make the original objective impossible while preserving the mission's underlying purpose.
Implement verification at three distinct layers—continuous…
Implement verification at three distinct layers—continuous signals, periodic samples, and infrequent structural audits—calibrated to the stakes, maturity, and track record of the delegation.
Verify outputs against predefined success criteria rather…
Verify outputs against predefined success criteria rather than monitoring methods, preserving delegate autonomy while maintaining accountability for results.
Build verification that is transparent to the delegate…
Build verification that is transparent to the delegate rather than hidden, making checkpoints explicit and predictable to maintain trust while ensuring accountability.
Extend trust incrementally based on accumulated evidence…
Extend trust incrementally based on accumulated evidence of reliability rather than as a fixed initial condition, tightening verification when errors surface and loosening it as performance stabilizes.
When delegating to tools, verify outputs more aggressively…
When delegating to tools, verify outputs more aggressively than when delegating to humans because tools lack judgment, cannot flag their own uncertainty, and produce errors with uniform confidence.
Treat every tool as a component of your extended cognitive…
Treat every tool as a component of your extended cognitive system rather than as an external convenience, selecting and configuring tools based on architectural criteria of reliability, maintainability, interoperability, and failure-mode transparency.
Maintain practice of capabilities you delegate to tools…
Maintain practice of capabilities you delegate to tools to prevent atrophy of the biological original, especially for capabilities whose loss would compromise judgment or effectiveness.
Design habits as contextual cue-routine pairs rather…
Design habits as contextual cue-routine pairs rather than motivation-dependent behaviors, making the context itself trigger the automatic behavior without requiring conscious decision.
Align habit rewards intrinsically to the behavior itself…
Align habit rewards intrinsically to the behavior itself rather than adding external incentives, since genuine automaticity requires satisfaction built into the completion rather than bolted-on consequences.
Verify habit automaticity by checking whether the behavior…
Verify habit automaticity by checking whether the behavior fires from context without conscious decision rather than checking whether the behavior occurs consistently, since consistency can be maintained through willpower without successful delegation.
Write documentation that answers specific questions…
Write documentation that answers specific questions with full decision context (constraints, alternatives, rationale) so the document can replace the author's presence entirely.
Force clear thinking by requiring complete sentences…
Force clear thinking by requiring complete sentences with explicit causal structure rather than bullet points that allow ideas to remain unexamined.
Create pre-committed decision rules for recurring…
Create pre-committed decision rules for recurring, low-to-moderate consequence situations to eliminate deliberation at every instance while preserving judgment for novel cases.
Schedule periodic rule audits to detect when changed context
Schedule periodic rule audits to detect when changed context has made previously effective rules obsolete or when rules are being used to avoid necessary decisions.
Maintain ability to explain what delegated systems are doing
Maintain ability to explain what delegated systems are doing in concrete terms—if you cannot articulate method, tradeoffs, and quality criteria, you have over-delegated.
Preserve the ability to perform delegated work at acceptable
Preserve the ability to perform delegated work at acceptable levels yourself—if the delegate's disappearance would leave you unable to function, delegation has become dependency.