--- title: "Do more AI agents mean fewer tokens? What one benchmark measured · Cork" description: "Do more AI coding agents use fewer tokens? Usually more. What Anthropic's docs and research report, and what Cork's 12-agent benchmark measured." canonical: "https://corkcli.com/blog/more-ai-agents-fewer-tokens-benchmark" last_updated: "2026-10-03" published: "2026-10-03" author: "Kevin Vera" --- # Do more AI agents mean fewer tokens? What one benchmark measured Do more AI coding agents use fewer tokens? Usually more. What Anthropic's docs and research report, and what Cork's 12-agent benchmark measured. By Kevin Vera · October 3, 2026 · [Guides](https://corkcli.com/blog/c/guides) · 6 min read No: adding AI agents almost always means more tokens in total, not fewer. Each agent is a separate model session with its own context window, and published figures from Anthropic put multi-agent systems at many times the token use of a single session. What coordination can change is how many of those tokens are wasted. In Cork's published benchmark (12 Claude Code agents at once on one repository, hard tasks), the same agents used 53% fewer tokens with Cork than without it. That is a measurement of coordination overhead among a fixed number of agents, not evidence that more agents are cheaper. ## What the published numbers say about multi-agent token cost The clearest public figures come from Anthropic. - **Multi-agent research.** In its June 2025 engineering write-up on its multi-agent research system, Anthropic reported that agents typically use about 4 times more tokens than chat interactions, and multi-agent systems about 15 times more than chats. The same post reported that a multi-agent setup (Claude Opus 4 leading Claude Sonnet 4 subagents) outperformed single-agent Claude Opus 4 by 90.2% on its internal research evaluation, and that token usage by itself explained 80% of the variance in performance on the BrowseComp evaluation. In other words, the extra tokens bought better answers on broad research tasks. - **[Claude Code agent teams](https://corkcli.com/blog/cork-vs-claude-code-agent-teams).** Claude Code's cost documentation says agent teams use approximately 7 times more tokens than standard sessions when teammates run in plan mode, that token usage is roughly proportional to team size, and that each teammate keeps consuming tokens until it exits. The agent teams docs recommend starting with 3 to 5 teammates. The pattern is consistent: more agents, more tokens. Whether that is worth it depends on whether the work parallelizes. Anthropic's post makes the same point: multi-agent systems excel at tasks with heavy parallelization, while domains that require all agents to share the same context or involve many dependencies between agents are not a good fit today, and "most coding tasks involve fewer truly parallelizable tasks than research." ## Where the tokens go when coding agents share a repo Coding is the hard case because parallel agents on one codebase are rarely independent. Tokens get spent on work that a single agent would not have done: - **Rediscovery.** Each agent searches and reads the same files to learn the same structure, because none of them knows what the others already found. - **Stale assumptions.** An agent writes code against a function that another agent renamed ten minutes ago, then spends turns diagnosing the failure. - **Landing retries.** When several agents push to one branch, most pushes are rejected as non-fast-forward, and each retry (fetch, rebase, rerun tests, push) costs model turns. - **Conflict repair.** Two agents change the same function in different directions, and someone, often a third agent, spends tokens reconciling them. None of this is useful work. It scales with the number of pairs of agents that can interfere, which grows faster than the number of agents. ## What the Cork benchmark measured Cork's benchmark, published at , ran 12 Claude Code agents at once on one repository with hard tasks, once with Cork and once without it. Agents with Cork used 53% fewer tokens than the same agents without it. How to read it: - **It compares coordination, not team size.** The number of agents was the same in both arms. The result says coordinated agents wasted fewer tokens than uncoordinated ones; it does not say 12 agents are cheaper than 1. - **It is one vendor's benchmark on one setup.** It was run by Cork on its chosen repository and tasks. Treat it as a data point and measure on your own codebase. - **Correctness is reported beside the tokens.** A token reduction that came with wrong answers would be meaningless; the study grades every task with hidden tests and reports 11 of 12 correct with Cork and 12 of 12 without. ## How to measure this on your own team You can run a small version of the same comparison: 1. **Pick tasks that overlap.** Independent tasks will not show coordination effects. Include at least one rename or signature change that other tasks depend on. 2. **Fix everything else.** Same model, same prompts, same number of agents, same starting commit. 3. **Grade on outcome.** Use tests that run on the final shared branch, not the agents' own reports. 4. **Count tokens per arm.** Claude Code's `/usage` shows a session's token usage by model, including cache reads and writes; organization analytics and OpenTelemetry export give per-user totals. Sum input, output and cache tokens for every agent, including subagents. 5. **Run more than once.** Agent runs vary; one run of each arm is an anecdote. If coordinated agents do not come out ahead on your tasks, your work may already be well partitioned, and you can scale agents with less overhead. ## Ways to cut wasted tokens with or without a tool - Partition tasks so agents own different files or packages. - Write specific prompts; vague prompts trigger broad scanning. - Give agents symbol navigation (language server integrations) instead of text search. - Brief agents at session start with recent changes via a hook. - Land through a script that rebases, tests and pushes, so retries happen outside the model. - Use cheaper models for routine agents and shut idle ones down. ## How Cork helps Cork aims at the wasted part of multi-agent token spend. Its hooks in Claude Code, Codex and Cursor work with no extra model calls: each agent is briefed on the exact signatures, callers and recent changes of the code it is about to touch, so it searches less; when another agent changes code yours depends on, your agent hears it on its next step; a [write gate](https://corkcli.com/blog/write-gate) holds an edit against code another agent is changing right now until your agent has their version; and `git push` becomes one step that rebases, reruns the tests and pushes. The `recap` tool reports what happened and what it cost. Cork is the [coordination layer](https://corkcli.com/blog/coordination-layer-ai-coding-agents) for multiplayer coding agents: it keeps Claude Code, Codex and Cursor agents, on one machine or across a team's machines, from overwriting each other on one codebase, and hands each agent what the others already learned. ```bash curl -fsSL https://corkcli.com/install.sh | sh ``` ## Frequently asked questions ### Do multi-agent systems use more tokens than a single agent? Yes. Anthropic reported multi-agent systems using about 15 times more tokens than chat interactions, and Claude Code's docs say agent team usage is roughly proportional to team size. ### Can coordination reduce token usage for parallel coding agents? In Cork's published benchmark of 12 Claude Code agents at once on one repository (hard tasks), agents with Cork used 53% fewer tokens than the same agents without it. Measure on your own tasks before relying on it. ### Is it worth running more agents if it costs more tokens? When tasks are independent and parallelize well, often yes; Anthropic found more tokens tracked better results on broad research tasks. For tightly coupled code, coordination overhead can eat much of the gain. ## Sources - - - - ## Related reading - [How to reduce Claude Code token usage](https://corkcli.com/blog/reduce-claude-code-token-usage): Cut Claude Code token usage with the official levers: /clear, /compact, /context, model and effort choice, a lean CLAUDE.md, subagents, hooks, fewer re-reads. - [How to review pull requests from several AI agents at once](https://corkcli.com/blog/review-pull-requests-from-multiple-ai-agents): Review pull requests from several AI coding agents at once: triage with gh, review for cross-PR conflicts, use CODEOWNERS, AI review passes and a merge queue. - [Codex hooks and Cursor hooks: how to use them for team coordination](https://corkcli.com/blog/codex-cursor-hooks-team-coordination): How Codex and Cursor hooks work as of October 2026, where the config lives, how to block or brief an agent, and how to use hooks to coordinate a team's agents. - [MCP servers for coordinating coding agents: what they are and how to use one](https://corkcli.com/blog/mcp-servers-coordinating-coding-agents): What an MCP server for coordinating coding agents does, how it differs from hooks, and how to connect one to Claude Code, Codex, Cursor and Gemini CLI. [All articles](https://corkcli.com/blog.md) · [Cork in Markdown](https://corkcli.com/index.md) · [llms.txt](https://corkcli.com/llms.txt) This page in HTML: https://corkcli.com/blog/more-ai-agents-fewer-tokens-benchmark