Subagents vs. skills: stop using subagents to hold instructions

A decode of @dexhorthy's subagent critique and Anthropic's fold-back guidance, for Claude Code users deciding between subagents and skills.

Picture a freelance full-stack developer who maintains three client apps in Claude Code. Over the summer she wrote a subagent for each thing she kept repeating: one that knows a client's React conventions, one that knows how that client wants API errors shaped, and one that writes tests their way. Now every feature fans out to three helpers. Each one starts cold, re-reads files the main session already opened, and hands back a summary that drops the detail she needed. The session looks busy and parallel. The diffs still need a lot of fixing.

That's the pattern @dexhorthy called out on X on October 24, 2026: "Subagent overuse and abuse has been a problem since mid 2025." He names the "biggest crime" as "conflating context management with instruction modules (these are quite orthogonal concerns)." His post points to an hour-long AI That Works episode with @vaibcode. Two days earlier, Anthropic published guidance arguing that many enterprise agents do better when subagents are folded back into one loop that uses skills. The working rule from both: use subagents to isolate context, and use skills to steer behavior.

What it is

This isn't a product launch. It's a design argument coming from two directions.

The AI That Works episode, published March 16, 2026, breaks down commands, skills, agents and sub-agents in a harness like Claude Code. According to its description, it covers removing irrelevant context "ASAP," why separating instruction modules from context isolation matters, why you shouldn't put custom instructions directly into your agents, and the risks of leaning on agent swarms and "lights off" software factories. We couldn't view the video, so its specific claims aren't verified here.

@dexhorthy's post gives the history. Before skills, subagents were the only way to manage a small set of different system prompts. Now, he writes, skills are the best method for instruction modules, meaning steering a behavior or process, and "in most harnesses" they can be invoked from a parent agent or a subagent.

Anthropic's guide (October 22, 2026) comes from its Applied AI team. It says that "across several enterprise deployments, a single agent with skills frequently outperformed both the one-prompt-for-everything design and the subagent design on quality, and often at a lower cost and latency per task." The approach it describes is progressive disclosure: the agent sees short descriptions of skills and tools, and their full content loads only when needed. Anthropic says it cut Claude Code's system prompt for more advanced models "by over 80%," partly through progressively disclosing skills and tools. Independent write-ups from Clauding and SmartScope summarize the same guide. Ccleaks adds that Anthropic recommends grouping tools by user intent, because tool choice tends to degrade past 30 to 50 tool definitions.

How this maps to the lede: a non-fork Claude Code subagent doesn't see your conversation history, skills you already invoked, or files the main session already read, according to Alex Cloudstar's guide on DEV. That's the cold start. Instructions held in a subagent are only available when that subagent runs. Instructions held in a skill can be loaded on demand wherever the work is happening.

Alternatives, each with a trade-off:

Who it's for

Pick this up if your .claude/agents/ folder is mostly agents that exist to hold "how we do X" instructions, or if you run a router, planner and domain-subagent stack on Claude Platform. Unrequested subagents can pile up quietly. Explore AI Together found 28 subagent files installed by a plugin, plus a rule telling Claude to use them "PROACTIVELY."

Skip the fold-back, and keep your subagents, when they are doing real context isolation. That means reading thirty files to answer one question, running disposable parallel tasks, verification, or separating permissions. Anthropic's guide keeps subagents in three cases: a single loop shows measurable degradation from context bloat, parallel tasks need fresh context windows that can be thrown away, or one agent struggles to pick tools or faces conflicting instructions in its system prompt (Anthropic).

Before and after

Before:

  1. The main session gets a task.
  2. It delegates to a "conventions" subagent, which starts without the conversation history and re-reads files.
  3. A summary comes back, and some detail is lost in the handoff. Anthropic's guide calls this losing "details in the back and forth."
  4. The main session delegates again for tests, with another cold start.
  5. You reconcile the summaries and fix the diff.

After:

  1. The main session gets a task and sees short skill descriptions.
  2. It loads the conventions skill and the testing skill into its own context only when they're relevant.
  3. It does the work with everything it has already read still in view.
  4. It sends only bulky, read-heavy or parallel work to a subagent, such as an Explore-style codebase search or a batch migration in worktrees.
  5. You review one coherent diff.

Try it

  1. Inventory. Subagent files live in .claude/agents/ (project) and ~/.claude/agents/ (user), plus a plugin's agents/ directory and managed settings (Claude Code docs). Check enabled plugins too.
  2. Tag each one. Is it there to isolate context or to hold instructions? Keep the isolation agents. The instruction agents are candidates to become skills. None of our sources show the skill file format, so check the Claude Code docs for that rather than relying on a guess here.
  3. Measure before you fold. SmartScope suggests building about 10 representative test cases from existing workflows, folding subagents in one at a time, and comparing scores (SmartScope). For Claude Platform agents, Anthropic offers a migration kit with a guided workflow and a demo on a mock system (Anthropic). Ccleaks names the repo single-agent-loop-kit and reports that its reference loop defaults to 25 turns, 400,000 tokens and 16 tool calls per batch (Ccleaks).
  4. Test without the built-ins. Setting CLAUDE_CODE_DISABLE_EXPLORE_PLAN_AGENTS=1 removes the built-in Explore and Plan subagents and requires Claude Code v2.1.198 or later. Denying the Agent tool in permissions.deny stops all delegation (Claude Code docs).
  5. Keep isolation where it pays off. For batched writes, Anthropic's power user tips show this frontmatter:
# .claude/agents/worktree-worker.md
---
name: worktree-worker
model: haiku
isolation: worktree
---

Where it breaks

Sources