References: Orchestration that survives contact
Source material
Section titled “Source material”This lesson’s orchestrator-workers definition, its subagent briefing elements, its parallelism figure, and its “Agents are stateful and errors compound” line are quoted verbatim from Anthropic’s public engineering posts, and the context-isolation sentence is quoted from the Claude Agent SDK documentation. Rosa’s research crew, the two decomposition tests, the failure-report format, and the rumor rule are Clawdemy’s own synthesis and examples.
- How we built our multi-agent research system, Anthropic engineering blog, June 2025. The production account behind most of this lesson: a lead agent that plans and spawns parallel search subagents, the briefing guidance (“Each subagent needs an objective, an output format, guidance on the tools and sources to use, and clear task boundaries”), the up-to-90-percent parallelism result, the token economics of multi-agent systems, the compounding-errors reality, and the dedicated citation stage that pins final claims to sources.
- Building Effective AI Agents, Anthropic engineering blog, December 19, 2024. Defines the orchestrator-workers workflow this lesson builds on: “a central LLM dynamically breaks down tasks, delegates them to worker LLMs, and synthesizes their results,” suited to “complex tasks where you can’t predict the subtasks needed.”
- Subagents in the SDK, Claude Agent SDK documentation. The mechanics under this lesson: defining subagents with a description, prompt, and tool list; context isolation (“Intermediate tool calls and results stay inside the subagent; only its final message returns to the parent”); what a subagent does and does not inherit; and running subagents in parallel.
Going deeper
Section titled “Going deeper”- Introduction to subagents, in the Anthropic Academy catalog. Free, self-paced official course on using and creating subagents in Claude Code to manage context and delegate tasks, a hands-on companion to this lesson’s design discipline.
- Claude Code subagents documentation, Anthropic. The filesystem-based way to define subagents as markdown files, including tool restrictions per agent, least privilege from lesson 4 applied at team scale.
On this site
Section titled “On this site”- Tools other agents can trust (lesson 4). The contract discipline this lesson promotes from tools to agents: descriptions for readers who cannot ask, errors an agent can act on, least privilege.
- Orchestration and shared state (Track 25). The read-a-real-system view of coordination: how a working multi-agent team actually passes state, before this lesson’s design-it-yourself view.
Adjacent topics
Section titled “Adjacent topics”- The next lesson in this track is reliability: escalation paths, human review, and the decisions an agent team must not make alone. The gap annotations from this lesson become its triggers.
- If the token economics of seams interested you most, the Building with Claude track’s lessons on prompt caching and context management are the cost-side deep dive.