hardAI Engineering

When should you split work across multiple agents instead of one?

Q: When should you split work across multiple agents instead of one?

**TL;DR.** Default to **one capable agent with good tools**. Split for three real reasons: **context isolation** (a subagent burns tokens exploring, returns a summary), **parallelism** (independent subtasks fan out), or **role/permission separation** (a reviewer that must not share the writer's context or credentials). Multi-agent adds real costs: lossy communication, orchestration failures, harder debugging. orchestrator clean main context <polygon class='d-a

947 views

Understand the problem

Subagents, orchestrators and the honest trade-offs vs one capable agent with good tools.

multi-agentorchestrationsubagentsarchitecture

Attempt it yourself

Sketch your approach before reading the solution — that's what interviews test.

Stuck? AI Nudge Available

Get a conceptual hint to guide your logic without spoiling the final implementation.

Study the solution

The solution is waiting

Give it an honest attempt first — then compare your thinking with the full walkthrough.

Read the code

Subagent as context compression

// orchestrator delegates exploration; only the summary enters its context
const findings = await runSubagent({
  goal: "Locate where invoice totals are computed and list every caller.",
  tools: [grep, readFile],                 // read-only, scoped
  budget: { steps: 30, tokens: 80_000 },
  returns: "a <=500-token structured summary with file:line references",
});

orchestrator.push(user("Exploration result:\n" + findings.summary));
// the subagent's 60k-token transcript is logged for debugging, not carried forward

Join the discussion

Discussion (0)

No responses yet. Be the first to share what you think.

What is semantic caching and when is it safe to use?

How do you evaluate and debug agent trajectories?