「worktreesは死んだ。クラウドエージェントこそ未来」を、47万行の遺跡で確かめるA hands-on experiment using Cursor Pro ($20/month) tested cloud agents' three…

A developer blog post in the "Legacy Archaeology" series has put a popular claim to the test: that Git worktrees are obsolete and that swarms of cloud agents represent the future of AI-assisted coding. To check whether that assertion holds up outside of greenfield demos, the author ran an experiment using Cursor Pro, the $20-per-month tier of the AI code editor, against a genuinely difficult target: roughly 470,000 lines of legacy Java built on Apache Struts, drawn from the Mifos microfinance platform. The result matters because most enthusiasm for autonomous coding agents is demonstrated on small, modern, easily runnable projects, and this test deliberately chooses the opposite conditions.

The experiment, described as the ninth entry in the series, frames the "cloud agent swarm" argument around three pillars: execution, parallelism, and judgment. Execution refers to an agent actually running code to confirm that its changes work. Parallelism refers to fanning many agents out across a problem at once, replacing the manual juggling of Git worktrees that developers otherwise use to keep multiple working copies of a branch side by side. Judgment refers to having agents review or arbitrate one another's output. The post measures each pillar against the aging codebase rather than accepting the original claim at face value.

According to the write-up, the execution pillar effectively collapsed. None of the verification agents could start Mifos, so the number of confirmations obtained by actually running the application was zero. This is a familiar problem with old enterprise stacks: frameworks such as Struts often depend on specific application servers, database schemas, build tooling, and configuration that are difficult to reproduce, and a cloud agent operating in an isolated sandbox is unlikely to reconstruct that environment on its own. Notably, the author reports that the agents did not fabricate results to cover the gap. Instead, the approach was partially salvaged through what the post calls "carve-out execution," isolating self-contained, dependency-free calculations and re-implementing them in Python or Java so they could be run and checked in isolation. That technique restored some ability to verify behavior, but only for narrow slices of logic that could be cleanly separated from the surrounding framework.

The parallelism pillar fared better, though the post says it stood up only halfway. The excerpt indicates that breadth was a decisive advantage, meaning many agents could cover a wide surface of the codebase simultaneously, which is precisely the strength that swarm advocates emphasize over sequential, worktree-based workflows. The framing that parallelism "half stood" suggests that breadth alone did not translate into complete success, likely because coverage without reliable execution leaves the correctness of individual changes unproven. The source excerpt cuts off before detailing the judgment pillar, so the full verdict on agent-versus-agent review is not captured here.

Some background helps explain why this test is meaningful. Cursor is an AI-centric code editor that integrates large language models into everyday editing, and its cloud or background agents can work on tasks asynchronously rather than only responding to inline prompts. Git worktrees, the tooling the original claim declares "dead," let a single repository check out several branches into separate directories at once, a common way to run parallel experiments without constantly switching branches. Mifos is an open-source platform aimed at financial inclusion, and its older components rest on Struts, a Java web framework whose classic versions are now largely superseded. Combining a large line count with a dated framework produces exactly the kind of codebase where modern automated tooling tends to struggle, which is what makes it a useful stress test.

The broader takeaway appears to be a tempering of the "swarm is the future" narrative rather than a rejection of it. The experiment suggests that the value of cloud agents on legacy systems is uneven: parallel breadth is real and useful, but automated runtime verification, arguably the most important pillar for trusting changes, can fail outright when the software cannot even be started. For teams weighing whether to lean on autonomous agents for maintaining old enterprise code, the results imply that human-designed scaffolding, such as carving out testable units, remains necessary. As a single hands-on account on one codebase, the findings are not definitive, but they offer a concrete, measured counterpoint to sweeping claims about the death of established workflows.