---
title: "Context engineering for multi-agent systems"
description: "In multi-agent systems the hard part is what each agent sees: give sub-agents a scoped brief, pass results by reference, and keep decisions in one thread."
image: "https://bombos.ai/assets/og-en-v2.png"
---

# How do you do context engineering for multi-agent systems?

Short answer

Context engineering for multi-agent systems is deciding what each agent sees and what it hands back. The working pattern from Anthropic, Cognition and Google: give each sub-agent a scoped brief with objective, output format and boundaries, let it work in its own clean window, return a short result or a reference to a stored artifact, and keep decisions that must agree with each other in one agent.

*Piet Baudoin · published November 17, 2026 · November 17, 2026*

## Two camps that agree more than they seem

In June 2025, one day apart, two well-known posts seemed to contradict each other. Cognition’s “Don’t Build Multi-Agents” by Walden Yan argued that parallel agents produce fragile systems. Anthropic’s “How we built our multi-agent research system” reported that a lead agent with parallel sub-agents beat a single agent by 90.2% on its internal research eval. Read closely, both are about the same thing: context.

| Source | Main point | What it means for context |
| --- | --- | --- |
| Cognition (June 12, 2025) | “Share context, and share full agent traces, not just individual messages” and “Actions carry implicit decisions, and conflicting decisions carry bad results” | Agents that make decisions which must fit together should share one thread |
| Anthropic (June 13, 2025) | Orchestrator plus parallel search sub-agents works for broad research; weaker where agents must share context or depend on each other | Sub-agents work when their tasks are separable and their output is a finding, not a decision |
| Google ADK (December 4, 2025) | “Scope by default: Every model call and sub-agent sees the minimum context required” | Context is a compiled view per call; handoffs are explicit |

Cognition’s example: split “build a Flappy Bird clone” into two sub-tasks, and one agent builds a Super Mario style background while the other builds a bird that does not match. Each did its piece; the pieces conflict. Anthropic’s research agents avoid this because each one searches and reports facts, and only the lead agent decides.

## The practices

1.  **Write a real brief per sub-agent.** Anthropic found that vague instructions led to duplicated work and gaps. Each sub-agent needs an objective, an output format, guidance on tools and sources, and clear task boundaries.
2.  **Give sub-agents a clean window.** The point of a sub-agent is often context control: it reads and searches in its own window, and the main agent gets a summary instead of all the raw tool output. Cognition notes that Claude Code’s sub-agents are usually only asked to answer a question, not to write code, and do not run in parallel with the main work.
3.  **Pass references, not copies.** To avoid what Anthropic calls a “game of telephone”, sub-agents store outputs in external systems and pass a lightweight reference back. Google’s ADK does the same with artifacts: “Large data lives in the artifact store, not the prompt.”
4.  **Keep the plan outside the window.** Anthropic’s lead agent saves its plan to memory, because past 200,000 tokens the context gets truncated and the plan must survive.
5.  **Label who did what.** When control moves to another agent, ADK recasts earlier assistant turns as narrative such as “\[For context\]: Agent B said…”. Without that, the new agent may believe it took those actions itself.
6.  **Choose shared or separate history on purpose.** Google’s whitepaper names both patterns: one unified log for tightly coupled work, or private histories that only exchange final outputs through agent-as-a-tool or the A2A protocol.
7.  **Budget for tokens.** Anthropic reports multi-agent systems use about 15 times the tokens of a chat, and that token usage alone explained 80% of performance variance on BrowseComp. Use them where the task is worth it.

## The common misunderstanding

The usual assumption is that more agents means more intelligence, so you split the task by role (planner, coder, tester) and let them talk. The evidence points elsewhere. Splitting helps when sub-tasks are independent and return facts. It hurts when sub-tasks carry implicit decisions that must agree, because each agent fills the gaps differently. The design question is not “how many agents” but “what does each one need to see, and who decides”.

## What this means for a team

The same rule applies to people with AI tools. When five colleagues each run their own assistant, they are a multi-agent system without a shared trace: each AI makes implicit decisions about tone, pricing or architecture with only its own user’s context. Cognition’s principle holds here too. Shared decisions need shared context, which means team knowledge and agreed rules stored once and loaded into every agent, not rebuilt in each chat. See [How do you set up context engineering for a team?](https://bombos.ai/en/questions/context-engineering/how-to-set-up-context-engineering-for-a-team/) and, for keeping long agent sessions healthy, [What is context compaction?](https://bombos.ai/en/questions/context-engineering/what-is-context-compaction/)

## Where Bombos fits

Bombos gives the agents of a team, whether that is Claude Code, Codex, Cursor, ChatGPT or your own orchestrator and its sub-agents, one shared memory to read from: team knowledge, learned rules and what is already in mail, documents and repos. Each agent can retrieve what its task needs, and you can see what an answer was based on. One person can try it first for their own agents.

## Related questions

Are multi-agent systems better than a single agent?

For broad, parallel research Anthropic measured a large gain: its multi-agent setup beat single-agent Claude Opus 4 by 90.2% on an internal research eval. For tasks where agents must share context or depend on each other, such as most coding, Anthropic and Cognition both advise caution.

Why do multi-agent systems use so many tokens?

Each agent gets its own context. Anthropic reports agents use about 4 times the tokens of a chat, and multi-agent systems about 15 times.

Should sub-agents see the full conversation history?

Usually not. Google's ADK team recommends "scope by default": every sub-agent sees the minimum context required and fetches more through tools. Cognition argues the opposite for agents that make decisions together: share full traces, or keep them in one thread.

What is the "game of telephone" problem?

Each handoff through the lead agent loses detail. Anthropic's fix is to let sub-agents write their output to external storage and pass back a lightweight reference.

## Sources

1.  [How we built our multi-agent research system (Anthropic, June 2025)](https://www.anthropic.com/engineering/multi-agent-research-system)
2.  [Don't Build Multi-Agents (Cognition, Walden Yan, June 2025)](https://cognition.com/blog/dont-build-multi-agents)
3.  [Architecting efficient context-aware multi-agent framework for production (Google Developers Blog, December 2025)](https://developers.googleblog.com/architecting-efficient-context-aware-multi-agent-framework-for-production/)
4.  [Context Engineering: Sessions & Memory (Google whitepaper on Kaggle, November 2025)](https://www.kaggle.com/whitepaper-context-engineering-sessions-and-memory)

## What does your AI forget?

1.  01You tell us which AI you use and where it keeps starting over.
2.  02We show what a shared memory changes there, on your own examples.
3.  03You keep the plan, with or without Bombos.

![Piet Baudoin](https://bombos.ai/assets/team/piet-klein.jpg)

**Piet Baudoin**You talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe

## More questions about Context engineering

-   [What is Context Engineering 2.0?](https://bombos.ai/en/questions/context-engineering/what-is-context-engineering-2-0/)
-   [How does LangChain approach context engineering?](https://bombos.ai/en/questions/context-engineering/how-does-langchain-approach-context-engineering/)
-   [What is agentic context engineering (ACE)?](https://bombos.ai/en/questions/context-engineering/what-is-agentic-context-engineering/)
-   [What is advanced context engineering for coding agents?](https://bombos.ai/en/questions/context-engineering/advanced-context-engineering-for-coding-agents/)
-   [What is the difference between context engineering and prompt engineering?](https://bombos.ai/en/questions/context-engineering/context-engineering-vs-prompt-engineering/)
-   [How do you set up context engineering for a team using AI tools?](https://bombos.ai/en/questions/context-engineering/how-to-set-up-context-engineering-for-a-team/)
-   [RAG vs context engineering: is RAG still needed?](https://bombos.ai/en/questions/context-engineering/rag-vs-context-engineering/)
-   [What did Manus learn about context engineering for AI agents?](https://bombos.ai/en/questions/context-engineering/what-did-manus-learn-about-context-engineering/)

[All questions about Context engineering →](https://bombos.ai/en/questions/context-engineering/)


---

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://bombos.ai/#organization","name":"Bombos","url":"https://bombos.ai/","logo":"https://bombos.ai/assets/bombos-bee.png","description":"Bombos is the memory layer under every AI your team uses. One shared memory for Claude, ChatGPT, Copilot, Cursor and your own agents, that learns how you work.","sameAs":["https://www.linkedin.com/company/bombos-ai","https://www.youtube.com/@truebombos","https://www.instagram.com/truebombos/","https://bsky.app/profile/bombostrue.bsky.social","https://www.threads.com/@truebombos","https://mastodon.social/@Bombos","https://www.tiktok.com/@truebombos","https://t.me/truebombos","https://www.pinterest.com/truebombos/","https://github.com/bombos-ai","https://linktr.ee/bombos.ai","https://bombosai.substack.com","https://hashnode.com/@bombosai","https://hackernoon.com/u/bombos","https://dev.to/bombos"],"email":"info@bombos.ai","areaServed":["NL","Europe"],"knowsLanguage":["nl","en"],"contactPoint":{"@type":"ContactPoint","email":"info@bombos.ai","contactType":"sales","url":"https://bombos.ai/en/contact/","availableLanguage":["nl","en"]}},{"@type":"WebSite","@id":"https://bombos.ai/#website","name":"Bombos","url":"https://bombos.ai/","inLanguage":["nl-NL","en"],"publisher":{"@id":"https://bombos.ai/#organization"}}]}

{"@context":"https://schema.org","@type":"Article","headline":"How do you do context engineering for multi-agent systems?","description":"In multi-agent systems the hard part is what each agent sees: give sub-agents a scoped brief, pass results by reference, and keep decisions in one thread.","abstract":"Context engineering for multi-agent systems is deciding what each agent sees and what it hands back. The working pattern from Anthropic, Cognition and Google: give each sub-agent a scoped brief with objective, output format and boundaries, let it work in its own clean window, return a short result or a reference to a stored artifact, and keep decisions that must agree with each other in one agent.","datePublished":"2026-11-17T00:00:00.000Z","dateModified":"2026-11-17T00:00:00.000Z","author":{"@type":"Person","name":"Piet Baudoin"},"publisher":{"@type":"Organization","@id":"https://bombos.ai/#organization","name":"Bombos","logo":{"@type":"ImageObject","url":"https://bombos.ai/assets/bombos-bee.png"}},"image":"https://bombos.ai/assets/og-en-v2.png","mainEntityOfPage":"https://bombos.ai/en/questions/context-engineering/context-engineering-for-multi-agent-systems/","inLanguage":"en-US","isPartOf":{"@type":"CollectionPage","name":"Context engineering","url":"https://bombos.ai/en/questions/context-engineering/"},"citation":[{"@type":"CreativeWork","name":"How we built our multi-agent research system (Anthropic, June 2025)","url":"https://www.anthropic.com/engineering/multi-agent-research-system"},{"@type":"CreativeWork","name":"Don't Build Multi-Agents (Cognition, Walden Yan, June 2025)","url":"https://cognition.com/blog/dont-build-multi-agents"},{"@type":"CreativeWork","name":"Architecting efficient context-aware multi-agent framework for production (Google Developers Blog, December 2025)","url":"https://developers.googleblog.com/architecting-efficient-context-aware-multi-agent-framework-for-production/"},{"@type":"CreativeWork","name":"Context Engineering: Sessions & Memory (Google whitepaper on Kaggle, November 2025)","url":"https://www.kaggle.com/whitepaper-context-engineering-sessions-and-memory"}]}

{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Bombos","item":"https://bombos.ai/en/"},{"@type":"ListItem","position":2,"name":"Questions","item":"https://bombos.ai/en/questions/"},{"@type":"ListItem","position":3,"name":"Context engineering","item":"https://bombos.ai/en/questions/context-engineering/"},{"@type":"ListItem","position":4,"name":"How do you do context engineering for multi-agent systems?","item":"https://bombos.ai/en/questions/context-engineering/context-engineering-for-multi-agent-systems/"}]}

{"@context":"https://schema.org","@type":"FAQPage","mainEntity":[{"@type":"Question","name":"Are multi-agent systems better than a single agent?","acceptedAnswer":{"@type":"Answer","text":"For broad, parallel research Anthropic measured a large gain: its multi-agent setup beat single-agent Claude Opus 4 by 90.2% on an internal research eval. For tasks where agents must share context or depend on each other, such as most coding, Anthropic and Cognition both advise caution."}},{"@type":"Question","name":"Why do multi-agent systems use so many tokens?","acceptedAnswer":{"@type":"Answer","text":"Each agent gets its own context. Anthropic reports agents use about 4 times the tokens of a chat, and multi-agent systems about 15 times."}},{"@type":"Question","name":"Should sub-agents see the full conversation history?","acceptedAnswer":{"@type":"Answer","text":"Usually not. Google's ADK team recommends \"scope by default\": every sub-agent sees the minimum context required and fetches more through tools. Cognition argues the opposite for agents that make decisions together: share full traces, or keep them in one thread."}},{"@type":"Question","name":"What is the \"game of telephone\" problem?","acceptedAnswer":{"@type":"Answer","text":"Each handoff through the lead agent loses detail. Anthropic's fix is to let sub-agents write their output to external storage and pass back a lightweight reference."}}]}
```
