Assess agent ecosystem health by checking three pair-level failures: conflicting outputs, throughput mismatches, and resource competition

Why This Is a Rule

Individual agents can be functioning perfectly while pair-level interactions produce system failures. Agent A outputs high-quality work. Agent B produces high-quality work. But A and B together produce conflict because their outputs contradict, or B can't consume A's output fast enough, or both compete for the same time block. Pair-level assessment catches failures that individual agent assessment misses.

Three pair-level failure modes cover the majority of interaction problems: Conflicting outputs: two agents produce recommendations or behaviors that contradict each other (health agent says rest, productivity agent says push through). Resolution: scope separation (Define non-overlapping agent scopes — scope collisions are architecture problems, not willpower failures) or priority hierarchy (Assign exactly one agent as accountable for each decision — consulted agents advise, but only one has final authority). Throughput mismatch: a producer agent generates output faster than a consumer agent can process it (planning agent produces tasks faster than execution agent can complete them), creating a growing backlog. Resolution: throttle the producer or increase consumer capacity. Resource competition: two agents claim the same limited resource (both need the same 2-hour morning block). Resolution: allocation mechanism (Match allocation mechanism to contention structure — priority queue for unequal importance, rotation for equal, time-slicing for concurrent needs).

Checking all agent pairs during health assessments is the systematic way to find these interactions before they produce system-level failures.

When This Fires

• During periodic agent system health reviews (monthly or quarterly)
• When the system "isn't working" despite individual agents functioning well — pair-level issues are likely
• When adding a new agent to an existing system — check its interactions with all existing agents
• Complements Define non-overlapping agent scopes — scope collisions are architecture problems, not willpower failures (scope design) and Resolve inter-agent conflicts with documented priority hierarchies — case-by-case deliberation defeats the purpose of automation (conflict resolution) with the diagnostic that identifies which pairs need intervention

Common Failure Mode

Assessing agents individually: "My writing agent is working. My exercise agent is working. My review agent is working." All true — but the writing agent and exercise agent compete for the same morning block, and the review agent's output backlog keeps growing because the writing agent produces faster than review can process. Pair-level assessment reveals these interactions.

The Protocol

(1) List all active agents. For each pair (N agents = N×(N-1)/2 pairs, but focus on pairs that share time, resources, or data flows): (2) Check for conflicting outputs: do these two agents ever produce contradictory recommendations or behaviors? If yes → resolve with scope separation or priority hierarchy. (3) Check for throughput mismatch: does one agent produce output that the other consumes? If yes, is the producer faster than the consumer? If yes → throttle producer, increase consumer capacity, or add a buffer. (4) Check for resource competition: do these agents claim the same resource (time, attention, energy)? If yes → apply an allocation mechanism (Match allocation mechanism to contention structure — priority queue for unequal importance, rotation for equal, time-slicing for concurrent needs). (5) For systems with many agents, prioritize pairs that share resources or data flows — these have the highest interaction-failure probability.