What a Claude Code Subagent Actually Costs: Measuring the ~436k-Token Fixed Overhead
Real pipeline measurements reveal that a Claude Code subagent carries a massive fixed token overhead before doing any actual work.
Stock photo for illustration only, not from the actual event
- Spawning a subagent in Claude Code incurs a fixed overhead of roughly 436,000 tokens.
- This overhead stems from initial context loading and cache writing, independent of the task.
- The embedded product content being reviewed accounted for only about 46,000 tokens per agent.
- Developers recommend merging overlapping reviewer agents to avoid paying the fixed cost twice.
Spawning a subagent in Claude Code feels free, but it certainly is not. Measured across a real review pipeline, the number that matters and is almost never discussed is that each subagent costs roughly 436,000 tokens in fixed overhead before performing any useful work.
This post explores the origin of that figure, how to reproduce the measurement on your own setup, and how it alters the strategy of splitting work between AI agents. The team runs a weekly review pipeline over a catalog of digital products featuring Markdown-heavy repositories, embedding each product's full content into a reviewer prompt to request structured findings.
Breaking down the transcript per turn, each agent carried about 436k tokens of overhead unrelated to the review itself, driven by the initial context load at spin-up and the cache write on its final turn. Meanwhile, the embedded product content—assumed by many to dominate the cost—was only about 46k tokens per agent, yielding a striking 9.5-to-1 ratio.
Stock photo for illustration only, not from the actual event
No special tooling is required for measurement since Claude Code writes full transcripts as JSONL files under ~/.claude/projects/<project-dir>/, where each assistant message records its token usage. The exact overhead figure varies depending on system prompts, MCP servers, and loaded skills, as every always-on tool schema contributes to the spin-up payload.
Recognizing the fixed overhead of AI subagents is vital for optimizing multi-agent architectures. As highlighted in this analysis, initialization payloads such as system instructions and tool definitions can drastically outweigh the actual task data. Designing efficient AI workflows requires balancing agent autonomy against the hidden token costs of context initialization.
To adapt to these findings, teams should merge reviewers with overlapping perspectives, spawn subagents strictly for true independence and parallelism rather than tidiness, and avoid starving active agents by supplying complete files and contexts since the payload itself is the inexpensive part.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment