Single-Agent vs Multi-Agent¶
When multi-agent architecture is justified, when it's overkill, and when it's a correctness requirement.
Decision Criteria (Report 2 §9)¶
| Criterion | Single Agent | Multi-Agent |
|---|---|---|
| Independent verification needed? | Not required | Required (correctness) |
| Task decomposition clear? | Yes, single context | Complex, needs specialisation |
| Concurrent perspectives needed? | No | Yes |
| Error correlation acceptable? | Yes | No (safety-critical) |
| Coordination overhead justified? | N/A | Must exceed single-agent value |
When Single-Agent is Preferred¶
- Well-bounded tasks within single-agent capability
- Structured task decomposition (Pattern 3 orchestrator-worker suffices)
- No independence requirement
- Simpler is better: less coordination overhead, easier auditability
- Cost and latency constraints
When Multi-Agent is Required¶
- Independent verification: Safety-critical roles where correlated errors defeat the purpose of redundancy → epistemic-independence with model-heterogeneity
- Concurrent specialist perspectives: Multiple domains must be assessed simultaneously during time-critical events
- Adversarial challenge: Productive disagreement requires genuine independence (same-model debate suppresses independent reasoning)
When Multi-Agent is Overkill¶
- Shift handover summaries (Pattern 7 single agent suffices)
- Document retrieval and synthesis
- Simple parameter monitoring
- Any task where a single agent with good tools produces adequate results
The Graded Approach¶
Report 4 Scenario 8 proposes adaptive architecture: single-agent during normal operations, escalating to multi-agent during transients. Match complexity to need.