# One person, 298 agents, 5 days 14.5 hours: a CRM rebuilt in Rust

<!--
title: "Study 9: one person, 298 Claude Code agents, 26 workflow runs, a CRM rebuilt in Rust · Cork"
description: "A field study, not a with/without benchmark. Over 5 days 14.5 hours, one person ran 26 multi-agent workflows (298 Claude Code agents) that grew a CRM's Rust server from 1,490 to 140,811 lines with 964 tests, from 65 typed prompts. Cork's hooks briefed 176 of the 182 surviving agent transcripts; its gate never had to hold a write, because every stream owned its own files. What broke was power, network, stalls and a usage limit. About $3.7k API-equivalent; nothing reached production."
canonical: "https://corkcli.com/benchmarks/study-9-298-agents-one-person"
last_updated: "2026-11-15"
-->

Every other study here runs the same tasks with Cork and without. This one doesn't: it is a week of real work, read back from its transcripts. One person ran Claude Code's multi-agent workflows for 5 days 14.5 hours (Oct 5–10, 2026) to replace Guava's CRM, a self-hosted Twenty, with a Rust API and a Next.js UI. Cork's hooks (then named Collide) were on the machine the whole time. The question here is not how much Cork saved, which a week with no control run can't answer, but what running 298 agents for a week looks like, and what Cork did in it.

298
:   agents in 26 workflow runs, from one person's 65 typed prompts

1,490 → 140,811
:   lines of Rust in the CRM server, with 964 tests (from 25)

176 of 182
:   surviving agent transcripts briefed by Cork; writes held: 0

≈ $3.7k
:   API-equivalent cost (an estimate; it ran on a subscription)

## What Cork did

In the second session, whose transcripts survive whole, 176 of 182 agent transcripts received Cork's briefing: 520 briefings in all, each the exact signatures, callers and recent changes of the code the agent was about to touch. The six without one were short: three were Wave 8 retries that never got a response from the API, and three ran only a handful of steps.

Cork's gate never held a write, and no agent was told of a teammate's change mid-step. That is because of how the work was planned, not a gap in Cork: every workstream built in its own git worktree, owned a fixed list of files, and a single integrator merged the streams in a set order. Integration found no textual conflicts where the study records it (the redesign's six streams), and workflows run side by side never conflicted. When the plan keeps agents apart, there is nothing for a gate to catch, and coordination comes down to two things Cork does supply: briefing each agent on the code it is about to change, and landing in order.

When agents share files, it is a different picture. In [Study 7](https://corkcli.com/benchmarks/study-7-scaling-to-sixteen-agents.md) and [Study 8](https://corkcli.com/benchmarks/study-8-sixteen-agents-a-regression.md), sixteen agents on vague tickets that overlap, rejected pushes and merge conflicts are where most of the tokens go without Cork.

Briefings did not stop agents looking. Of 22,761 tool calls in the second session, 19,306 were shell commands, and about 11,662 of those began with grep, find, cat, ls, head, tail, sed -n, wc or a git log, show, diff or status, on top of 2,244 file reads. With no run of the same work without Cork, this study can't say how many of those the briefings saved; Studies 1 to 8 measure that.

## How the work was run

- **Waves.** A plan of 272 capabilities (9 done, 35 partial, 228 missing) became waves of four to seven workstreams. Each stream was built in its own worktree, reviewed by an agent told to find fault (ship, fix then ship, or rework), fixed, then merged in a fixed order and verified.
- **Pipelines, not barriers.** From the redesign on, each stream ran build, review and fix on its own clock, so a review started as soon as its build finished.
- **Skeptics as switches.** Before a risky step, agents were told to refute its safety claim, with "refuted" as the default when unsure, and the script skipped the step on a refutation. One rehearsal touching production data was refused twice, each time for a real hazard, and passed the third time with a narrower procedure.
- **Clean room.** Twenty is AGPL, so exactly one agent was allowed to look at it. It wrote a 6,400-word spec of the interface in words and numbers, and everyone else worked from the spec; a critic still found and removed four leaks.
- **The person's part.** 65 typed prompts, about 6,700 characters, more than half of them on the first day. They set direction in a line ("the server should be in rust"), answered gates, corrected wrong assumptions about the stack, and kept the laptop awake. The longest stretch with agents working and no message from the person was 21 hours, four waves chained back to back.

## What broke

None of it was agents overwriting each other.

- **Internet outages.** Wave 1's resume stalled on all six attempts while the connection was down; Wave 3's last six agents died mid-step. Both were closed by hand-written follow-up runs.
- **Power and sleep.** The laptop slept, was closed in a bag, lost network and rebooted. The person raised power, sleep or the lid at least seven times; it was the main source of failures.
- **Stalls.** Builders, integrators and fixers stalled and were retried; the retries found the earlier attempt's work on disk and carried it forward. Wave 7's integrator stalled three times, and integration took 7 h 17 min before a fourth attempt fixed the race underneath.
- **A guard that matched itself.** The main loop's own check for running builds matched its own command line and deadlocked two streams. Another agent in the run diagnosed it in about 2.5 minutes and messaged the main loop, which fixed the guard for the next wave.
- **The weekly usage limit.** The run ended mid-merge when the account hit its weekly limit: Wave 8 stopped with four of six streams merged on a local branch, and the transcript ends with no hand-off.

What the reviewing agents caught before merge (86 findings in the redesign runs alone) included a database role meant to be read-only that could read synced email bodies; a removed member who could be re-admitted with their old role; staging set to send real mail through production's mail account; a 1.8 KB crafted spreadsheet that crashed the server; and a bulk-edit shortcut that cleared a field on every selected row.

## Tokens and cost

The second session, deduplicated per API response: 19.5M output tokens, 6.32B cache reads and 135.8M cache writes. About 98% of input was cache reads: long agents that run many shell commands re-read a large context on every turn.

At Claude Opus 5.5's list prices that is about $2,369; the first session's own harness record says $1,345; together about $3,714. It is an estimate of what the same usage would cost through the API. The work ran on a subscription with a weekly limit, so it is not what was paid, and the limit, not the money, is what ended the week.

The first session's transcript was deleted by Claude Code's 30-day cleanup, so its runs are known from a table the second session printed while it still existed, and from git; their token counts are gone. Per-run figures (agents, start and end, and for the second session API responses, tool calls, output and cache-read tokens) are in the data: https://corkcli.com/benchmarks/study-9-298-agents-one-person/data.json

## Outcome

Waves 0 to 7, the redesign and its integration reached the development branch and its staging deployment; Wave 8 was finished by another session on Oct 16. Not cut over: nothing reached main or production. The person held promotion during the week, and on Nov 1 decided the old CRM keeps running for now. The CRM server ended the window at 140,811 lines of Rust with 964 tests, and the web side at 110,504 lines of TypeScript with 2,030 test cases and 58 end-to-end specs, counted in the code at that commit; whether they all passed is what the agents reported.

## What this study can't tell you

- How much Cork saved. There is no run of the same work without it.
- Anything about the first session's tokens: its transcript is gone; only git, its prompts and a cost record survive.
- Whether the tests the agents reported all passed: test counts and CI results are what the agents said; only commits and diffs were checked independently.
- Why some stalls happened, or exactly when the laptop rebooted: inferred, not recorded.

[All benchmarks](https://corkcli.com/benchmarks.md) · [Cork in Markdown](https://corkcli.com/index.md) · [llms.txt](https://corkcli.com/llms.txt)

This page in HTML: https://corkcli.com/benchmarks/study-9-298-agents-one-person
