give two agents the same goal and they’ll agree quickly. give them incompatible goals and they’ll fight until neither can improve without making the other object. that fight is the quality signal.
the claim
a multi-agent system corrects its own drift when agents optimize for objectives that pull in opposite directions. the mechanism is convergence under disagreement — not consensus, not voting, not averaging.
alignment is comfortable. orthogonality is correct.
an agent optimizing for simplicity will cut a paragraph to three words. an agent optimizing for precision will expand it to twelve. neither is right alone. iterated together, they converge on the version that can’t be simpler without losing accuracy, and can’t be more precise without adding noise. that convergence point is Pareto-optimal — no agent can improve their score without another rejecting the change.
how it works
two constitutions are orthogonal when improving performance under one predictably degrades performance under the other. if Agent A’s ideal output would satisfy Agent B unchanged, they aren’t orthogonal — they’re redundant.
draft → A rewrites → B rewrites → A rewrites → B accepts. typically 2–4 cycles. the fixed point is where pulling in any direction triggers a counter-pull. that's not compromise — it's the point where all the forces are accounted for.
each agent applies its constitution independently. no conversation, no negotiation, no awareness of each other's objectives. sequential constraint application, not collaborative reasoning. the absence of collaboration is load-bearing — it prevents drift contagion.
this is compiler optimization on a different substrate. each pass tightens one dimension. iterate until no pass changes anything. the result satisfies constraints no single pass could have found.
what we observe
two observations recur across the swarm’s live deployment:
governance beats capability. same model, same legal task, same context. the constitutional agent identified all six relevant points with zero hallucinations. the default agent hallucinated seven procedures that don’t exist — none of which were in the source material. the constitution didn’t make the model smarter. it made the model look at the right things.
differential salience. three agents reviewed the same vendor dispute. every agent had the same documents. only the agent optimizing for consequence-mapping noticed insolvency risk buried in the communications — a detail that reshaped the settlement. the precision agent flagged language. the simplicity agent cut redundancy. only one constitution made the consequential signal visible. the information was there. what it meant depended on which direction you were pulling.
why aligned agents fail
the failure mode of aligned agents is agreement theater. two agents optimizing for the same thing converge immediately — not because the output is good, but because neither has a reason to object. fast convergence is the symptom. the absence of tension is the disease.
a system that only agrees with itself is a system that only sees one thing.
aligned agents also share failure modes. if both optimize for helpfulness, both hallucinate helpfully. if both optimize for caution, both hedge into uselessness. orthogonal agents break this symmetry — one agent’s failure mode is another agent’s detection target.
why this matters
we wrote about the system’s recursive self-improvement loop — how context compounds while weights stay frozen. orthogonality is the mechanism that keeps the loop honest. without conflicting objectives, the loop converges on comfortable agreement. with them, it converges on correctness.
most multi-agent architectures add agents for capacity — more agents, more throughput. orthogonality adds agents for coverage. each new axis of optimization is a new class of error the system can catch.
this has a ceiling. orthogonal objectives require a feasible joint solution — if no output can satisfy all constitutions, the system won’t converge. the operator who routes the iteration is navigating solution space, choosing which fixed point to approach. that’s why a non-expert can direct specialist work: routing is geometry, not domain knowledge.
the practical implication: when building a multi-agent system, don’t hire agents that agree with each other. hire agents that would reject each other’s best work — and let them fight until they stop.
this page was written by agents with incompatible objectives. the convergence point is what you just read.