Subagent Context Passing, Coordinator Prompts and Session Forking

1.Subagents start fresh: the prompt is the handoff

Each subagent starts a new conversation. It loads its own system prompt and project-level context such as CLAUDE.md. It does not get the coordinator's message history. So any fact, file path, decision or earlier finding the subagent needs has to be written into the prompt the coordinator passes to it. Isolation works in the other direction too: the subagent's intermediate tool calls and results stay inside it, and only its final message goes back to the coordinator. That is why a subagent can read dozens of files without filling up the coordinator's context.

Separate invocations do not share memory either. Each query() call starts a new session. To continue a subagent's earlier work, you must resume the same session and refer to the subagent by its agent ID. If you skip that, a later call has no access to what the earlier subagent did.

Sources12

2.Passing complete findings, with their attribution

A synthesis subagent cannot see what the search subagents found, so the coordinator has to put those findings into its prompt. A subagent returns only a short summary to its parent. When a later stage needs earlier results, the coordinator should pass the complete findings: the web-search results *and* the document-analysis outputs. A paraphrase of them is not enough. Anthropic's Research system shows the same handoff. When research is finished, the lead agent passes all findings to a CitationAgent. That agent matches claims to specific source locations so that every claim stays attributed.

The exam guide adds a practice for these handoffs: use structured data that keeps each finding's content separate from its metadata, such as source URL, document name and page number. That way the next agent can still say where each claim came from. The sources here back the goal, which is attribution that survives the trip to the citation step. They do not define a schema or format for this. If prior outputs are too large to copy through the coordinator, the engineering post describes another option: subagents save their work to external storage and pass back lightweight references.

Sources23

3.Coordinator prompts: goals and quality criteria, not scripts

Open-ended research can't be planned step by step in advance, because each finding changes what to look at next. Anthropic's lead agent therefore does not hand subagents a fixed procedure. Each delegation gives an objective, an output format, guidance on which tools and sources to use, and clear task boundaries. The prompts aim to build good heuristics instead of rigid rules: break the problem down, judge source quality, and adjust the approach as new information arrives.

Too little guidance also fails. Short instructions like "research the semiconductor shortage" made subagents misread the task or repeat each other's searches. What helps is to be specific about the goal and the quality bar, and leave the steps to the subagent. The subagent's own system prompt, the prompt field of its AgentDefinition, sets its role and behaviour. The coordinator's per-task prompt says what this particular run should produce.

Sources23

4.Forking a session to explore other approaches

Sometimes you want to try two different approaches from the same analysis. Resuming a session adds to its existing history. Forking creates a new session that starts with a copy of that history, so you can try something else while the original stays as it was. Both use resume. A fork also sets fork_session=True (forkSession in TypeScript).

        async for message in query(
            prompt="Instead of JWT, outline how OAuth2 would work for the auth module",
            options=ClaudeAgentOptions(
                resume=session_id,
                fork_session=True,
                max_turns=5,
            ),
        ):

After the fork runs, resuming the original session_id carries on with the JWT approach as if the fork never happened. To branch from an earlier point instead of the end of the session, the Python SDK has resume_session_at. Use it with resume, and usually with fork_session, to load the conversation only up to a chosen message UUID.

Sources45

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.A subagent inherits the coordinator's conversation, so a short prompt that points back to earlier discussion is enough.Why is that wrong?

   Each subagent starts a new conversation and never sees the parent's turns, so every piece of context it needs must be written into its prompt.

   Covered in Subagents start fresh: the prompt is the handoff

2. 2.Exploring an alternative by resuming the same session keeps the original analysis clean.Why is that wrong?

   A plain resume adds to the original session. Only a fork creates a new session ID and leaves the original unchanged.

   Covered in Forking a session to explore other approaches