MAS Decision Gate

Meta-Rule

If a single agent can do the job, do not use MAS. Multi-agent systems are organizational scaling tools, not capability multipliers by default. Research shows 80% of AI projects fail due to premature architectural complexity.

12 Factor Agents Perspective (Factor 10)

The 12 Factor Agents framework reinforces this principle:

Factor 10: Small Focused Agents

Smaller focused prompts with controlled context always beat long autonomous runs.

This applies at two levels:

1. Single vs Multi-Agent: Start with single agent
2. Within Multi-Agent: Each agent should be small and focused

The Progression:

Level 0: Deterministic workflow (no agent)
    ↓ Only if judgment needed
Level 1: Single focused agent
    ↓ Only if tools needed
Level 2: Single agent with tools
    ↓ Only if verification critical
Level 3: Minimal MAS (Planner→Executor→Verifier)
    ↓ Only if multiple domains
Level 4: Full MAS (when justified by evidence)

Only advance levels when evidence supports it. Most tasks belong at Level 0-2.

Decision Criteria

Use MAS only if at least one is true:

1. Natural decomposition: Tasks split into semi-independent roles
2. Parallel benefit: Concurrent reasoning materially reduces latency (≥40%)
3. Distinct world models: Agents need different knowledge bases or incentives
4. Internal verification: Long-horizon work requires checks and balances

If none apply → build a single agent with structured tools.

Quantitative Thresholds

Deploy Single-Agent When:

Deploy Multi-Agent When:

Decision Questions

To determine whether MAS is appropriate, answer these questions:

Question 1: Can a single agent complete this task?

• If yes with reasonable quality → use single agent
• If struggling with scope/quality → consider MAS

Question 2: What distinct expertise areas are needed?

• Count domains requiring specialized knowledge
• 1-2 domains → single agent with tools
• 3+ domains → MAS may be justified

Question 3: Are subtasks truly independent?

• Can work proceed in parallel without dependencies?
• Yes → MAS provides latency benefit
• No → MAS adds coordination overhead without benefit

Question 4: Is internal verification critical?

• Would self-checking be insufficient?
• Do outputs need adversarial review?
• Yes → MAS with separate verifier agent

Question 5: What is the failure cost?

• Low-stakes task → prefer simplicity (single agent)
• High-stakes task → MAS verification may justify complexity

Simplicity Test

Before building any agent system, answer these sanity-check questions:

Core Questions

1. Could this just be a deterministic workflow or cron job?

   • If yes → use traditional automation, not agents

2. Where does uncertainty or judgment actually exist?

   • If nowhere → scripted workflow is sufficient
   • If bounded → single agent with tools
   • If distributed across domains → MAS may be justified

3. What would happen if the agent vanished tomorrow—could you survive?

   • If operations stop → high value, proceed carefully
   • If minor inconvenience → question the investment

4. What's the simplest version that would provide value?

   • Build that first, then add complexity only when evidence supports it

12 Factor Simplicity Questions

In addition to the core questions, apply these 12 Factor Agents checks:

5. Could you own the control flow with code? (Factor 8)

   • If yes → code + single agent likely suffices
   • If no → MAS may help distribute complexity
   • Key insight: Code-controlled DAG beats LLM-controlled DAG

6. Can state be modeled as (state, event) → new_state? (Factor 12)

   • If yes → cleaner architecture possible with reducers
   • If no → complexity is intrinsic, MAS may help
   • Key insight: Stateless reducers enable debugging and replay

7. Is context building well-understood? (Factor 3)

   • If yes → single agent with explicit context
   • If no → need to understand context before adding agents
   • Key insight: Context engineering is the core of agent quality

8. At what level does the task belong? (Factor 10)

   • Level 0: Deterministic (script/workflow)
   • Level 1-2: Single agent (with or without tools)
   • Level 3-4: Multi-agent (only if justified)
   • Key insight: Start simple, add agents only when evidence supports it

Red Flags (Agent May Be Overkill)

Green Flags (Agent Justified)

Simplicity Test Output

Document simplicity assessment:

## Simplicity Assessment

**Task**: [Brief description]

**Deterministic alternative?**: [Yes/No - what would it look like?]
**Where is judgment needed?**: [Specific points]
**If agent vanished?**: [Impact assessment]
**Minimum viable version**: [Description]

**Conclusion**: [Proceed with agent / Use simpler alternative]

Common Anti-Patterns

Anti-Pattern: MAS for Capability

Wrong: "Multiple agents will be smarter than one"

Reality: Coordination overhead often exceeds capability gains. ChatDev shows 25% correctness, 60-87% failure rates across frameworks.

Anti-Pattern: Premature Decomposition

Wrong: "Let's split this into 5 agents for better organization"

Reality: Every agent boundary introduces failure points (specification, alignment, verification). Start simple, add agents only when evidence supports it.

Anti-Pattern: Personality-Based Splitting

Wrong: "Creative agent, analytical agent, careful agent"

Reality: Split by functional orthogonality, not personality. Planner, Executor, Verifier—not "smart" vs "creative."

Specialist vs Generalist

2026 consensus strongly favors specialists:

Why specialists win:

• 40-60% fewer tokens for domain tasks
• Higher accuracy in specialized domains
• Clearer audit trails and governance
• Reduced computational waste

Architecture pattern: Specialist agents orchestrated by a coordinator that handles delegation.

Decision Output Format

After analysis, document the decision:

## MAS Decision

**Task**: [Brief description]

**Decision**: [Single-Agent / Multi-Agent]

**Rationale**:
- Domain count: [X] domains
- Parallel benefit: [Yes/No - expected %]
- Verification need: [Low/Medium/High]
- Failure cost: [Low/Medium/High]

**If Multi-Agent, justify each agent**:
- Agent 1: [Role] - [Why separate agent needed]
- Agent 2: [Role] - [Why separate agent needed]

Additional Resources

Reference Files

For detailed decision frameworks and evidence:

• references/evidence.md - Research data supporting thresholds
• references/decision-tree.md - Step-by-step decision flowchart
• ../agent-specification/references/twelve-factor-agents.md - Quick reference for all 12 factors

Related Skills

After deciding on MAS, use:

• agent-specification - For writing proper agent specs (Factors 1, 2, 4, 7)
• coordination-patterns - For choosing architecture (Factors 3, 5/6, 8, 12)
• production-readiness - For cost/observability planning (Factors 9, 11)