The orchestrator's context vs. the workers' context

Why this matters

In the common orchestrator-workers pattern, the orchestrator holds the big picture (the plan, progress, which workers ran) while each worker holds only its narrow sub-task context. Conflating these — giving workers the full plan or letting the orchestrator accumulate every worker's raw output — defeats the pattern. The orchestrator's context should stay high-level and bounded (a plan + summaries), and workers should be stateless-ish and focused. Keeping these two context levels distinct is what lets the pattern scale to many workers without the orchestrator's window exploding.

Demo

The demo separates orchestrator state (plan + per-task status + worker summaries) from worker state (just the sub-task + its inputs). Workers return summaries, not raw output, so the orchestrator's context grows slowly.

Try it yourself

1. Confirm the orchestrator stores summaries, not raw worker output — its window stays small as workers multiply.
2. Have a worker return raw output instead and watch the orchestrator's context grow with each step.
3. Add 20 workers and verify the orchestrator's context scales with summary count, not total work volume.
4. Give a worker the full plan and reason about why that's usually unnecessary coupling.

Prompt your AI

Use these three in order. Each builds on the one before.

1. Basics & terminology

In the orchestrator-workers agent pattern, how should context differ between the orchestrator and the workers?

2. Why it works (the mechanism)

Explain why workers should return compact summaries (not raw output) to the orchestrator, and how that keeps the orchestrator's context bounded as worker count grows.

3. Advanced — application & what's next

Design the context model for an orchestrator coordinating 50 parallel workers: what the orchestrator tracks, what workers return, and how to avoid the orchestrator's window becoming the bottleneck.

References

• Anthropic — Building effective agents
• Anthropic — Effective context engineering for agents

Chat about this lesson