Multi-Agent Context Isolation

Chapter 6 solved "how does a single agent run a long task without blowing up context". Production hits another scale — a main agent dispatching sub-agents. The choice between shared, isolated, or partially shared context decides whether the system can scale.

Three Multi-Agent Context Topologies

1. Shared Context — One LLM Runs the Whole Thing

Every step runs in the same messages array.

messages = [
  user: "调研 Kubernetes 三个最流行的 ingress controller 各自优劣"
  assistant: 我先列三个候选 → ingress-nginx, traefik, contour
  tool_use: WebFetch ingress-nginx docs
  tool_result: ...
  tool_use: WebFetch traefik docs
  tool_result: ...
  ... (累积 50K context)
  assistant: 综合分析后, ingress-nginx 优在生态成熟, traefik 优在...
]

Pro: easiest to implement (no state management); zero information loss.

Con: context grows linearly, the Lost in the Middle problem bites around step 30; every step pays the accumulated token cost.

Fits: under 10 steps with strong dependencies between steps.

2. Isolated Context — Sub-Agents Fully Independent

The sub-agent gets a fresh context — just task description + needed reference material. After running, it returns a final answer + short summary, not the process.

主 agent context:
  user: "调研 K8s 三个 ingress controller"
  assistant: 我会派 3 个 sub-agent 并行调研
  tool_use: spawn_subagent(name="research_ingress_nginx", task="...")
  tool_result: { summary: "ingress-nginx 优在生态成熟...", refs: [...] }
  tool_use: spawn_subagent(name="research_traefik", task="...")
  tool_result: { summary: "traefik 优在配置灵活...", refs: [...] }
  ...
  assistant: 综合三个 sub-agent 的总结...

# 每个 sub-agent 自己的 context（独立）：
  user: "调研 ingress-nginx：架构、性能、生态、坑"
  tool_use: WebFetch ingress-nginx docs
  ... (自己累积 30K context, 完成后丢弃)
  assistant: 给主 agent 一段 500 字摘要

Pro: main agent context stays small; sub-agents can run in parallel; one blowing up doesn't affect others.

Con: lossy handoff; main can't see intermediate reasoning, has to trust the result; needs spawn / coordinate infrastructure.

Fits: sub-tasks independent and parallelizable; main agent needs to scale past 30 steps.

3. Partial Sharing — Summary Plus Key Raw Data

Sub-agents return a summary, plus the key supporting evidence verbatim. Main agent gets summary + a few key quotes and can verify the sub-agent.

Anthropic's Multi-agent Research System blog pattern — main is Lead Researcher, sub-agents are Search Subagents. Handoff includes source URL + key quotes for verification.

Cost: handoff payload 5-10× larger than pure summary, but trust goes way up.

Anthropic Multi-Agent Research System — Official Case

Anthropic's 2025-04 blog describes the multi-agent research feature they built into Claude.ai:

Architecture:

LeadResearcher (主)
  └─ 拆解用户 query → 决定派几个 sub-agent
  └─ 派 SearchSubagent 1: "找 X 主题的最新 paper"
  └─ 派 SearchSubagent 2: "找 Y 公司的官方文档"
  └─ 派 SearchSubagent 3: "找用户评测和实际案例"
  └─ 收集所有 sub-agent 摘要 → 综合写最终 report

Anthropic's engineering takeaways (direct quotes):

• "Sub-agent context can't be shared — parallel sub-agents not knowing what others do actually avoids duplicate work and group think"
• "Optimal sub-agent count is 3-5; past 5, lead coordination cost outweighs the gain"
• "Sub-agents handing raw search results to lead is wrong — they have to distill into findings first"
• "The whole system uses 4× more tokens than a single-agent baseline, but the quality lift far outweighs the token cost"

Claude Code's Agent Tool — Sub-Agent in Production

Claude Code's built-in Agent tool is the sub-agent pattern (this course was very likely written by Claude Code — main agent dispatches an Explore subagent to look up prompt-master configs. The main agent doesn't read the 200-line config raw, just the structured summary returned).

Config:

• subagent_type — different sub-agents get different tool sets (Explore can only read, code-architect can design but not edit)
• isolation: "worktree" — high-blast-radius tasks dispatch to a sub-agent in a separate git worktree, then merge or discard
• Description and Prompt — sub-agent only sees its prompt + own tool list at startup

"Isolated context + lossy handoff" productized — main delegates, protects its own context.

JR Real Case: Isolation Across omni-report's 17 Routines

JR Academy's omni-report runs 17 independent routines (AI Visibility / Competitor Weekly / Marketing Topics / Growth Playbook / Daily Jobs ×4 / various daily reports). Each is its own cron job — independent context, git commit, Notion sync.

Why not merge into a single "omni-report master agent"?

• Context isolation — Competitor Weekly's 50K context shouldn't pollute Daily Jobs
• Failure isolation — one crashing doesn't stop the other 16
• Observability — each monitored separately
• Re-runnable — any can be triggered alone (the 13th routine's idle-timeout retry only affected that one)

Cross-routine flow: async via git commits. Marketing Topics runs Monday and writes to marketing-topics/$DATE.md. Growth Playbook runs Tuesday and Reads last week's report into its own context. "Filesystem as sub-agent handoff channel" — no message queue, git is enough.

Shared / Isolated / Partial — Trade-off

JR's internal rule: under 10 steps single LLM. 10-30 partial sharing. 30+ isolated sub-agent. The curve Anthropic verified building their own research system.

Takeaway

Multi-agent isn't "split as fine as possible". Optimal sub-agent count is 3-5 — past 5, coordination cost eats the gain. Sub-agents have to distill findings before handing back; raw data still blows up the main agent. Anthropic's own research system uses 4× tokens but the quality lift far outweighs cost — proving "more tokens but layered" beats "fewer tokens but tangled".

---

References

1. Anthropic. (2025-04-15). How we built our multi-agent research system — Lead + Search subagent architecture and the 3-5 sub-agent rule of thumb.
2. Anthropic. Claude Code documentation — Agent tool — sub-agent type / isolation / handoff implementation.
3. Anthropic. (2024-12-20). Building Effective Agents — original orchestrator-workers pattern.
4. LangGraph. Multi-agent supervisor pattern — open-source equivalent.
5. AutoGen. GitHub — Microsoft's multi-agent framework, comparative reference.

Production case: JR Academy omni-report — 17 independent routines using git commits for async cross-agent handoff, filesystem as sub-agent communication channel.