Specialist Agents or Specialist Prompts? The Cheaper Half of Multi-Agent

"We need a security-review agent, a performance agent, and a style agent." Often the right answer is one agent, three prompts, and no handoffs — which costs a fraction and loses nothing that mattered.

The confusion is between two things that get called the same thing: giving the model a specialized role, and running a specialized process with its own context and lifecycle. Only the second is multi-agent, and only the second carries multi-agent costs.

What a separate agent actually buys

Three things, and none of them is expertise:

A separate context. The specialist doesn't see — and doesn't pay for — everything the parent accumulated. For a task that must read a large document, or where the parent's history is irrelevant, this is real and substantial.

A separate tool and permission set. The security reviewer can hold credentials the style checker doesn't. This is a genuine isolation boundary and the most defensible reason to split.

Independent execution. Separate agents can run in parallel, retry independently, and fail without taking the others down.

Notice what's absent: nothing about a separate agent makes the model better at security review. The expertise comes from the prompt, and a prompt is free.

What separate prompts buy

Everything about focus, and nothing about isolation.

for lens in [security, performance, style]:
    findings += model(diff, system=PROMPTS[lens])

Three focused passes over the same input, no handoff, no serialization boundary, shared retrieval, one place to add a fourth lens. If the passes need the same context and the same tools, this is the multi-perspective system — the agent framing adds ceremony around it.

⚠️ And the passes here are genuinely independent in a way that helps: each sees the raw diff rather than a summary of another agent's findings, which is exactly what a chained multi-agent version would degrade.

The decision

Ask what each specialist needs to see and touch.

• Same input, same tools, different focus → separate prompts. One process, N passes, possibly parallel.
• Different context needs — one must read something large the others don't need → separate agents, so the bulk stays out of the parent context.
• Different permissions — one needs credentials the others shouldn't hold → separate agents, for isolation rather than expertise.
• Different lifecycles — one runs for an hour, one for a second; one retries, one must not → separate agents.
• Genuinely parallel with independent failure → separate agents, though separate prompts can run concurrently too.

The heuristic: if you'd struggle to say what the sub-agent's context contains that the parent's doesn't, you want a prompt.

The hybrid worth knowing

Frequently the right structure is separate prompts for the analysis and a separate agent for one specific step that needs isolation.

A code-review system: three prompt passes over the diff, in one process, sharing the fetched files. Then, for anything flagged as a potential security issue, a separate agent with the credentials to query the vulnerability database and read the deployment config — things the review pass shouldn't hold.

The split follows the permission boundary, not the topic boundary. That's the general shape: split where authority or context genuinely differs, not where the subject matter has a different name.

🔍 A test for existing systems

For each agent in your design, answer:

1. What's in its context that isn't in its parent's? (If nothing: it's a prompt.)
2. What can it do that its parent can't, or shouldn't? (If nothing: it's a prompt.)
3. What does it return, and what did it discard to produce that? (If the parent needed the discarded part: the split is hurting.)

Any agent failing all three is pure overhead — an extra call, a lossy handoff, and a component to maintain, in exchange for a role description you could have put in a prompt.

The takeaway

Multi-agent is a mechanism for separating context, permissions, and lifecycles. It is not a mechanism for adding expertise — prompts do that, without handoffs, without extra calls, and without discarding the evidence at every boundary. Reach for a second agent when you can name what its context or its authority must exclude. Otherwise you want a second prompt, and you'll get better results for less.