# Navigating Claude Code: Subagents Done Right > [!info] Ported from an automated deep-read — not Efimov's words except inside `> [!quote]` blocks. > Sourced 2026-08-16 from the Readwise deep-read pipeline's analysis of this HackerNoon article. Product-version claims (agent names, nesting limits, model inheritance) are release snapshots — verify against current Claude Code docs before treating them as fact. > [!question] Lingering Questions > - What evidence must a subagent return for the parent session to actually verify its work, rather than just trust a confident summary? > - When does handoff compression (a subagent's terse report) remove essential nuance the parent needed? > - How should conflicting findings from two subagents get resolved — whose report wins? > [!tip] Article Ideas > - Turn each of the article's three recommendations (isolate research, restrict tools/model, specify return shape) into a bounded, measured experiment on a real WordPress contribution task — token count, elapsed time, tool errors, human corrections. > [!link] Cross Reference > - [[3. Resources/Permanent Notes/Subagent Configuration Inherits From the Parent With No Override]] — a harness-level constraint (Codex-confirmed) that this article's "restrict tools and choose a model suited to the subtask" advice assumes you *can* do, and in at least one harness currently cannot > - [[3. Resources/Permanent Notes/A Smart Orchestrator Should Delegate to Cheaper Executors]] — the cost-routing half of the same delegation pattern > - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] > [!quote] Vocabulary > - *context isolation* — running a subtask in a separate context window so its intermediate exploration (file reads, failed attempts, verbose output) never enters the parent session's context ## Bib - **Source**: [Navigating Claude Code: Subagents Done Right](https://hackernoon.com/navigating-claude-code-subagents-done-right) — Oleg Efimov, HackerNoon, part of a ten-article Claude Code workflow series - **Reader doc**: [[3. Resources/Readwise/Articles/Navigating Claude Code Subagents Done Right]] ## Author's thesis Subagents protect the main session from verbose intermediate work only when three conditions hold together: the task is well scoped, the tools and model are restricted to fit the subtask, and the subagent is told to return a short, structured report rather than a narrative summary. Skip any one of the three and the isolation either does nothing or actively wastes context. ## Key arguments (≤3) 1. **Delegate independent research and high-volume analysis into isolated contexts** — the parent session sees the summary, not the journey; that's the entire value of isolation. 2. **Restrict tools and choose a model suited to the subtask** — a subagent with full tool access and a top-tier model defeats the cost/focus rationale for delegating in the first place. 3. **Specify the return shape** — a verbose narrative report cancels out the context savings that isolation was supposed to buy. ## Claim atoms ### Isolation only pays off when the parent never sees the journey > [!quote] ID:rw_01kz7stxbwdfd6xpjq1zrdcmhx > The main session sees the summary, not the journey. Where/why: the article's core delegation argument — this is the mechanism, not just a benefit, of subagent isolation. ### Restricting tools is what makes isolation meaningful > [!quote] ID:rw_01kz7sty0t67rh8aqpb8wwsdz6 > That isolation is the entire point. Where/why: without restricting tools/model to the subtask, "isolation" is just extra overhead with no payoff. ### Verbose subagent output erases the savings isolation was supposed to create > [!quote] ID:rw_01kz7styzcpmfv04ctv9e2yqw0 > Verbose output cancels out the context savings. Where/why: the return-shape requirement — a subagent that reports narratively defeats its own purpose. ## Permanent candidates - **"Subagent delegation only pays off when scope and report shape are both fixed"** — the strongest generalizable claim: isolation alone doesn't save context; it's the combination of a bounded task and a bounded report that does. Promoted below. ## Processing checklist - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] - [x] Create Permanent note for the scope+report-shape claim - [ ] Verify current Claude Code docs on model inheritance and nesting limits before citing them elsewhere ## Links - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — subagent context-isolation economics as a routing/cost decision - [[3. Resources/Permanent Notes/Subagent Delegation Only Pays Off When Scope and Report Shape Are Both Fixed]] — the new Permanent note this source generated - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — applies-in: Joins the thread on building and orchestrating agent systems.