ONE BRAIN OR MANY? DESIGN FOR HOW WORK REALLY GETS DONE

PARADOX

Multi-agent looks like power. In most enterprises, one accountable brain wins on speed, cost, and clarity.

---

WHEN A SINGLETON WINS

You have one primary objective, one main dataset, and a clear decision loop.
A single reasoning core plans, acts, and learns without cross-agent chatter.

• •

  Price governance at quote time.
  One agent reads customer context, product cost, contract terms, competitive signals, and returns a price with an explanation and guardrails.
  KPI: quote cycle time, realized margin, override rate.

• •

  Recovery and root cause.
  One agent triages failed orders or returns, pulls traces, proposes fixes, writes PRs or tickets, closes the loop.
  KPI: mean time to resolution, repeat-failure rate.

• •

  Assortment rationalization.
  One agent clusters SKUs, reconciles attributes, maps to a reference taxonomy, and outputs a rationalized hierarchy with audit trail.
  KPI: catalog coverage, duplicate reduction, explainability.

• •

  Rebate eligibility and accruals.
  One agent interprets program logic, validates transactions, flags exceptions, and posts accruals.
  KPI: leakage reduction, audit findings, close speed.

• •

  Sales assist for renewals.
  One agent scores risk, drafts outreach, and proposes save-offers tied to profitability bands.
  KPI: NRR, discount-to-save ratio, rep adoption.

---

WHEN MULTI-AGENT IS JUSTIFIED

You face non-stationary environments with separable skills, partial observability, and parallelizable work where coordination overhead is paid back in result quality.

• •

  Supply disruption response.
  Planner agent hypothesizes impact, Data agent ingests signals across vendors and lanes, Optimizer agent simulates allocations, Policy agent checks compliance, Writer agent drafts customer comms.
  KPI: service level preserved vs baseline, expedite cost avoided, time-to-playbook.

• •

  Technical document extraction at scale.
  Crawler agent fetches PDFs, Vision agent extracts tables and diagrams, Normalizer agent standardizes units, Verifier agent runs spot checks, Loader agent updates the graph.
  KPI: doc throughput, extraction accuracy, refresh latency.

• •

  Procurement triage for tail spend.
  Classifier agent routes intents, Bidder agent runs quick quotes, Risk agent screens vendors, Settlement agent validates terms, Coach agent briefs the buyer.
  KPI: cycle time, savings captured, policy violations prevented.

• •

  Integrated promo planning.
  Demand agent forecasts scenarios, Margin agent simulates price ladders, Inventory agent checks constraints, Contracts agent validates terms, Orchestrator reconciles tradeoffs.
  KPI: promo ROI, stockouts averted, contribution vs plan.

---

HYBRID PATTERN THAT WORKS

Hierarchy beats swarm. Use one accountable brain to plan, assign, reconcile, and learn. Let specialist agents execute bounded skills.

• •Orchestrator owns the goal, memory, and evaluation.
• •Specialists own tools: retrieval, math, simulation, extraction, messaging.
• •Shared store holds facts, decisions, critiques, and outcomes for replay.

---

ANTI-PATTERNS TO AVOID

Multi-agent is not a license to create a committee in code.

• •Specialists with overlapping mandates that chase different KPIs.
• •Chatty peers without a shared memory or arbitration.
• •Agents that can mutate objectives at runtime.
• •No post-decision learning tied to realized outcomes.

---

DECISION MATRIX

Pick the simplest architecture that closes the loop.

• •Choose a singleton if the task maps to one objective, one loop, and one primary toolchain.
• •Choose hybrid MAS if you need parallel skills that benefit from modularity, but outcomes must reconcile to one objective.
• •Avoid flat MAS when governance, cost control, or auditability are non-negotiable.

---

PROOFS THAT EXECUTIVES CARE ABOUT

Tie agents to business levers, not novelty.

• •Decision velocity: time from signal to action to measurable outcome.
• •Alignment: fraction of decisions within margin, service, and risk bands.
• •Cost curve: tokens, tool calls, and human touches per closed loop.
• •Learning rate: variance reduction after each iteration at the same volume.
• •Explainability: reproducible trace from input to action to result.

---

BUILD ORDER

Ship intelligence like a product, not a science fair.

• •Phase 1: Singleton that closes the loop with evals and memory.
• •Phase 2: Add specialists where clear bottlenecks exist.
• •Phase 3: Introduce an orchestrator policy that reconciles conflicts.
• •Phase 4: Automate counterfactuals and A/Bs to improve the policy.

---

BOTTOM LINE

Start with one brain that earns the right to delegate.
Scale hands only after you’ve built the mind.