One agent or subagents? When to fold them into one loop

What Anthropic found

In comparisons across several enterprise deployments, Anthropic says a single agent with skills frequently beat both the one-giant-prompt design and the subagent design on quality, often at lower cost and latency per task (source). A few specifics from the article:

• One customer that tested its multi-agent harness against a single agent loop saw roughly 2x faster runtime with the single loop.
• Subagents pass each other summaries, and details get lost at every handoff. In one loop, the model that writes the plan is the one carrying it out, so it can change the plan mid-task.
• Progressive disclosure fixes the old reason for splitting: the agent sees short skill and tool descriptions and loads the full text only when it needs it. Anthropic says it cut Claude Code's system prompt for more advanced models by over 80%, partly this way.
• Tool choice tends to degrade once you pass 30 to 50 tool definitions in the system prompt, so group tools by what the user is trying to do, not one tool per API.

When to keep a subagent

Anthropic is clear that subagents still have their place. Keep one when:

• a single loop measurably gets worse because its context is too full,
• several tasks need to run in parallel with fresh context you can throw away afterwards, or
• a single agent struggles to pick the right tools, or its instructions conflict.

If you do split, split by context, not by job title: give each subagent work that needs little from the rest of the system and returns a short result. Independent research paths, separate components with clean interfaces, and black-box verification are the good candidates.

How to fold one in without breaking things

The article's advice fits in one line: build a small end-to-end eval, fold in one subagent, and keep it as a subagent only if the folded version scores worse, then repeat. In practice for a solo builder:

1. Write 10 test tasks. Anthropic says even 10 golden examples is enough to start. Include every chain of steps that runs together and every rule the agent must not break.
2. Run each task more than once. In Anthropic's own test on a sample support agent, folding one subagent in broke a process that had been passing, and the break showed up in only one run out of two.
3. Fold one subagent. Turn its prompt into a skill (the steps and judgment calls) and its API calls into tools.
4. Compare which tasks flipped, not just the total score. Keep the fold only if nothing important broke.

Anthropic also published a migration kit with a demo you can try on a mock system first (single-agent-loop-kit on GitHub).

What this means for a vibe coder

If your setup has a "planner" agent, a "coder" agent and a "tester" agent passing notes to each other, try one agent with a clear rules file and a few skills instead, and keep a separate reviewer only for checking finished work with fresh eyes. That's the black-box verification case Anthropic still recommends. Our make your coding agent prove it's done checklist covers that reviewer step, and one shared rules file for Claude Code, Codex and Cursor covers the rules file.

Get the rest of the loop free

Our free Checkpoint Loop gives a single agent one check command that auto-detects npm, pytest, Cargo and Go, blocks obvious secrets in your diff and commits only on green.

Get the free Checkpoint Loop