---
title: "Study 2: ten AI coding agents on one repo, token cost measured · Cork"
description: "Ten Claude Code agents with Cork and ten without, same task, same model, same run: 240K vs 1,262K tokens per agent (5.27× fewer), cost per agent falling from 545K to 240K as agents joined. All 20 correct; every transcript checkable."
canonical: "https://corkcli.com/benchmarks/study-2-ten-agents-one-repo"
last_updated: "2026-09-26"
---

# Ten AI agents, one repo: 5.3× fewer tokens per agent

In a measured run of twenty Claude Code agents on the same task, the ten on Cork used 240K tokens each and the ten searching the repository themselves used 1,262K. With Cork, the cost per agent fell from 545K to 240K as agents joined. Every agent in both groups landed the change correctly.

## Tokens per agent

Bars are to scale. Agent 1 on Cork pays for discovery (545K); agents 2 to 10 inherit it. The agents without Cork range from 779K to 1,952K, with no trend.

Running cost per agent

## Every agent

Tokens are input, cache writes, cache reads and output, as the provider reported them. Cost is the model's list price for that usage. Totals: 2,395K tokens and $1.69 with Cork; 12,623K tokens and $4.91 without.

| Agent | Cork tokens | Turns | Cost | Without tokens | Turns | Cost |
| --- | --- | --- | --- | --- | --- | --- |
| 1 | 545,482 | 17 | $0.370 | 1,144,045 | 43 | $0.527 |
| 2 | 186,827 | 4 | $0.136 | 778,965 | 45 | $0.388 |
| 3 | 190,632 | 4 | $0.139 | 845,898 | 45 | $0.420 |
| 4 | 192,780 | 4 | $0.141 | 833,486 | 37 | $0.396 |
| 5 | 194,519 | 4 | $0.144 | 1,525,178 | 31 | $0.502 |
| 6 | 195,460 | 4 | $0.146 | 1,545,650 | 29 | $0.529 |
| 7 | 195,956 | 4 | $0.145 | 1,343,785 | 29 | $0.462 |
| 8 | 247,379 | 5 | $0.158 | 843,274 | 41 | $0.464 |
| 9 | 248,544 | 5 | $0.161 | 1,951,686 | 37 | $0.639 |
| 10 | 197,908 | 4 | $0.147 | 1,810,573 | 35 | $0.582 |

## What passed from one agent to the next

Agent 1 finds it

17 turns. It read the layers, found where the value lives in each, and made the change as typed operations saved as a recipe, verified by the repository's own check.

The server keeps it

The recipe is one row on Cork's ledger. It records the transformation, not the code, so it applies to any module of the same shape.

Agents 2–10 are handed it

Before each later agent said a word, the prompt hook put a briefing in its context naming the recipe. Each replayed it in one command, made its one body edit, and finished in four or five turns.

No agent messaged another in this run. The work moved through the server: facts in the briefing, a verified recipe to replay, and a check that graded each write. That is the mechanism that makes each additional agent cheaper.

## Atomic claims under load

Two agents creating the same migration number is a collision no code graph can see: the files differ, so nothing overlaps until both land. Cork's `claim` tool hands out the next number in one database statement. 500 claims fired at once from 50 agent sessions came back as 500 distinct numbers, 1 to 500 with no gaps, in 1.33 seconds; one claim on an idle server takes 77.2 ms (p50).

## How the numbers can be checked

- **Tokens.** Each total is the provider's own usage report. Summing each session transcript message by message gives exactly the same number for all 20 agents.
- **What happened.** Cork's server keeps a hash-chained ledger: each row carries the hash of the one before, so nothing can be edited or removed without breaking every hash after it. The 828 rows of this run verify end to end.
- **Correctness.** The repository's own check passes on both final trees. The agents' summaries were not used.

## What this study does not show

- Agents ran one after another, each on its own module. Four agents editing overlapping code at once is in Study 1.
- One run per condition. Study 1 measured the same task at 6.2× in Round 11, and the ratio moves with conditions like the next item.
- Claude Code's own per-session overhead grew by about 7,000 tokens per message since Study 1, for both groups. A fixed cost added to both sides narrows the ratio most for the shorter Cork sessions, and we believe it accounts for most of the difference from Study 1's 6.2×.
- Adding agents raises the total; what falls is the cost of each agent. Splitting one task across ten agents does not beat one agent on tokens.
- In dollars the gap is smaller than in tokens, because most of a long session is billed at the cache-read rate.

[All studies](https://corkcli.com/benchmarks)[Connect your agents](https://corkcli.com/docs)

This page in HTML: https://corkcli.com/benchmarks/study-2-ten-agents-one-repo