A single context window cannot sustain expert depth across incompatible reasoning stances. Give each phase its own specialist — architecture to one agent, review to another, documentation to a third — and the coherence cascade does the rest: superior early decisions automatically improve all downstream work.
A generalist agent handling a multi-step task produces surface-level coverage at each step: technically correct but lacking depth. Architecture gets hedged rather than decisive; review gets polite rather than rigorous; documentation gets comprehensive rather than clear. The agent spreads attention across incompatible reasoning modes and commits fully to none.
A single agent or generalist persona prompt handles a multi-step task — software architecture, research synthesis, document production, or similar — that requires different types of reasoning at each phase. Architecture phase requires generative, expansive thinking. Review phase requires critical, constrained thinking. Documentation phase requires translational thinking that takes prior decisions as fixed. The agent produces surface-level coverage at each step: technically correct but lacking depth.
Assign a distinct, specialized persona to each task phase. Each persona has constrained focus, inherits context from previous personas via explicit pass-forward, and contributes depth within its domain rather than coverage across all domains. The pass-forward is explicit and complete: each persona reads all prior output before acting.
The “coherence cascade” follows from this structure: when each persona reads its predecessors’ complete output, superior early decisions automatically improve all downstream work. A specialized persona cannot accidentally override earlier choices outside its constrained domain. This creates traceability that a generalist prompt cannot.
Supervisor-subordinate architecture works at scale: a supervisor agent breaks down work and delegates to specialist agents with narrower roles, smaller tool sets, and clearer permissions. The supervisor maintains the whole-task view; specialists maintain depth within their domain.
Limit the pipeline length based on task type. Parallelizable tasks benefit from multi-agent coordination (centralized coordination improved performance by 80.9% over single-agent on parallelizable work). Sequential reasoning tasks degrade with each added agent. Tool-heavy tasks pay a 2-6x efficiency penalty in multi-agent systems.
Expert-depth output per phase, with zero terminology inconsistencies compared to frequent drift with generalist prompts. Failures localize to the phase where they occurred, enabling targeted repair.
Upfront cost: the task structure must be analyzed before personas can be designed. Phases and their reasoning stances must be explicit.
Sequential dependency slows parallel work. Coordination costs increase with agent count. The empirical saturation threshold at approximately 45% single-agent accuracy means: once a single agent can already handle the task reasonably well, adding specialized agents yields diminishing or negative returns. Specialization gains are largest where single-agent performance is lowest.
Sagar Mandal’s ten-persona Knowledge Store Generator demonstrates the pattern at concrete scale: Product Architect, Product Strategist, Systems Architect, Quality Engineer, Domain Designer, Spec Writer, UX Designer, Test Architect, Reference Librarian, and a final Product Architect verification pass. Each persona constrains attention to its domain while inheriting the full context chain.
Google/MIT research on scaling agent systems found that on parallelizable tasks, centralized coordination improved performance by 80.9% over single-agent baseline. On tasks requiring strict sequential reasoning, every multi-agent variant tested degraded performance by 39-70%. Tool-heavy tasks suffer a 2-6x efficiency penalty in multi-agent systems. Independent agents can amplify errors up to 17x when mistakes propagate unchecked. The empirical saturation threshold exists at approximately 45% single-agent accuracy.
The Groundskeeper-Gardener commission architecture in this garden instantiates the pattern: the Groundskeeper holds architectural scope while the Gardener holds the local-patch view. Neither attempts the other’s function.
extracted_from::[[Persona and Agent Personalities]]↑