Multi-Agent Systems (MAS)

Multi-agent systems split work across specialized roles (for example: planner, researcher, executor, reviewer).

This is useful when one "general agent" becomes too slow, too expensive, or too error-prone for complex workflows.

How it works (architecture)

1. Orchestrator receives the task.
2. Task is decomposed into subtasks.
3. Specialist agents execute in sequence or parallel.
4. Reviewer/critic agent validates outputs.
5. Orchestrator merges final result.

Typical role split

• Planner: turns request into task graph
• Researcher: gathers evidence and sources
• Builder/Executor: performs implementation steps
• Reviewer: quality, policy, and consistency checks

Best use cases

• Complex workflows
• Parallel subtasks
• Separation of responsibilities
• Long processes requiring review loops
• Cross-functional operations (ops + legal + finance + product)

Trade-offs

• More orchestration overhead
• Harder debugging and evaluation
• Cost can rise quickly without strict routing

Real-world company and service examples

• OpenAI ecosystem: many teams implement planner + tool-executor + critic patterns using API tool calling.
  Approximate API range (model-dependent): about $0.20/$1.25 to $2.50/$15 per 1M input/output tokens for common standard tiers.
• Anthropic ecosystem: multi-step Claude workflows using role-specialized prompts and tool pipelines.
  Approximate API range: often $1/$5 to $3/$15 per 1M input/output tokens for common tiers.
• Enterprise process automation suites: orchestrate specialized agents for support, sales ops, reporting, and compliance checks.
  Typical pricing model: seat/subscription + usage; often $20-$150+ per user/month plus API costs.

Cost management strategy

• Route simple subtasks to low-cost models.
• Reserve premium models only for hard reasoning or final review.
• Cache intermediate results.
• Set max rounds per agent (to prevent loops).

Reliability strategy

• Add explicit contracts between agents (input/output schema).
• Require citations from researcher agents.
• Add reviewer veto rules for policy and factual errors.
• Keep full execution traces for debugging.

When to choose MAS

Choose MAS when: - A single-agent pipeline fails on quality or maintainability. - You need parallel workstreams with clear ownership. - You need audit trails and role-level controls.

Avoid MAS when: - The task is short and deterministic (use reflex or model-based reflex). - Team cannot maintain orchestration and evaluation infrastructure.