Single Agent or Multi-Agent System? Patterns, Handoffs and Failure Modes

Evidence check: 25 October 2026. Product details reflect the referenced documentation observed at this cutoff; exact plans, prices, regional eligibility and unverified release dates are not claimed. This is an editorial guide, not a hands-on test.

Begin with one accountable controller

One agent can classify a request, call a search tool, draft an answer and request approval. This architecture has fewer moving pieces, less context replication and a clearer audit trail. Create specialist agents only when you have independently measurable subproblems, distinct credentials, isolated datasets or different task ownership. A “researcher,” “critic” and “writer” talking to one another may add model calls without adding independent evidence.

Common orchestration patterns

• Router: a controller sends a request to one of several specialists; it must validate the selected destination.
• Sequential: one stage produces structured output consumed by the next; schema checks stop error propagation.
• Parallel: independent read-only subtasks execute concurrently, then a deterministic merger resolves conflicts.
• Supervisor: a parent delegates bounded tasks to workers and applies an overall stopping rule.
• Graph workflow: explicit states and conditional branches handle approvals, failures and recovery.

Handoffs need scoped data

Passing a conversation to another agent may expose user history and retrieved secrets that the recipient does not require. Send only task-specific context, record the handoff and restrict tool privileges. OpenAI Agents SDK supports handoffs and warns that input metadata is not a replacement for authorization. Microsoft Agent Framework differentiates dynamic agents from explicit workflows; CrewAI distinguishes Crews from Flows for a similar engineering choice.

Failure modes

A specialist can hallucinate a finding, and a downstream agent may treat it as verified. Two workers can write to the same record. A supervisor may fail to notice partial results. Multi-agent systems can also enter escalation loops or exceed usage limits. Use structured contracts with confidence or evidence fields, bounded fan-out, conflict resolution and human intervention when stages disagree.

Evaluation protocol

Compare multi-agent and single-agent baselines on the same held-out tasks. Count externally verified completion, error rate, latency, model/tool usage and approval burden. A more intricate architecture earns its place only if a real improvement survives those controls.

Recommended default

Use deterministic workflow infrastructure for sequencing and transaction management. Insert the minimum number of agents necessary for reasoning-heavy subtasks. Keep the user’s authority and the system’s policy at the top, not delegated to whichever agent speaks last.

Four coordination choices

Use a router when questions belong to clearly different tools or specialists. Use a planner-and-executor only when plans can be inspected and constrained. Use a supervisor-and-workers graph when parallel subtasks are materially independent. Use a fixed pipeline when the handoffs are known in advance. Do not call a multi-step function chain “multi-agent” unless separate model-driven decision-makers really exist. Handoffs should define input shape, output schema, timeout and allowed tools.

An evidence-merging exercise

Ask two agents to independently summarize an official migration guide, then combine conclusions. Require every statement to carry a source identifier and explicit uncertainty. If the agents disagree, create a contradiction record for a human reviewer instead of accepting the more confident version. Measure whether added agents actually reduce factual errors enough to justify more latency, token use and coordination complexity. Cross-agent prompt injection and accidental disclosure are failure modes unique to shared contexts and delegated permissions.