Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableHabr AI→ original

Claude Code harness worked for one person but broke for colleagues: 5 steps to portability

A harness is the set of skills, MCP connections, subagents and checks around an AI agent such as Claude Code. An author at Directum ran his for a year without failures, but it broke when colleagues tried it: the scripts still contained paths from his machine, and a clean virtual machine had no Python or Node.js. He describes five steps to remove himself from the setup. Paths are now resolved from the repository root, and the installer checks the environment on its own. The takeaway: test installation on a clean machine before the first colleague does.

AI-processed from Habr AI; edited by Hamidun News
Claude Code harness worked for one person but broke for colleagues: 5 steps to portability
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The review-and-fix loop is capped at three attempts, after which a human is called in.
Key metric #2A subscription is considerably cheaper than API tokens for the author's tasks, and no server is required.
Key metric #3Evals, or test scenarios, catch skill regressions after changes.

Many people have already built a personal harness around an AI agent: they've watched short videos, added a pile of skills, connected MCP, and ended up with a "harnessed" workspace. A harness is everything that surrounds the model: tools, instructions, knowledge base, subagents, and checks. While one person uses it, it works.

The author of this piece, a specialist at Directum who promotes the Directum RX platform, found out what happens next. His harness ran flawlessly for a year, but once he handed it to colleagues, it broke: its main component turned out to be the author himself, and you can't install the author through an installer. Below he describes five steps to remove himself from the setup.

Some of the solutions already work, others are still being refined, and he himself calls the text personal experience rather than a guide to building an AI team.

What exactly was being handed over

The author receives requirements from clients and needs to quickly show a prototype of how to implement them on the platform. The Directum RX development environment (Development Studio) has a JSON-RPC API, which is connected to Claude Code via MCP. The agent creates entities, properties, and method stubs through the same mechanism a developer uses in the studio window, then writes the business logic in C#. Packages are published to the test stand with the standard Deployment Tool, and the result is verified through the API and the web interface.

The work is divided among subagents: the product owner prepares requirements and acceptance criteria, the architect designs the solution, the developer writes code and tests, the tester checks scenarios, and three reviewers examine code, security, and architectural compliance in parallel. The artifacts of each stage are stored in a .pipeline folder, and a script manages the transitions: it checks mandatory outputs and saves state, so after a break the coordinator resumes from the same point.

There are two check-and-fix loops: one for the solution design and one for code. Each has a limit of three attempts. Without a limit, subagents can spend a long time enthusiastically fixing one another while burning through the subscription limits. If the third attempt fails, the coordinator calls a human and saves the bug description, check results, and the history of attempts.

Shared server or local installation

The author considered two options. The first is a shared service with sandboxes (such as GoClaw or OpenHands), where the user gets a preconfigured environment with the required product version. For a single team with one environment, this is convenient. But the harness had to serve several teams with different products, platform versions, and solution sets. One environment for everyone can't be assembled, and supporting multiple variants would require a separate team.

Therefore he chose a local installation: each person clones the harness from a separate repository on the internal GitLab and works under their own subscription. For the author's tasks, a subscription is considerably cheaper than paying for tokens through the API, and a shared service doesn't need to be deployed from day one. Teams connect gradually and see whether the harness helps. The cost of this choice is that dependencies, paths, access rights, and tool versions differ across machines.

Shared, project, and personal data are stored separately. General rules go into the harness repository, project requirements live next to the project's code, and personal settings and credentials stay local and never reach Git. While the author was the only user, the scripts had paths to his machine hardcoded. Now the harness looks for its own files relative to the repository root, and the paths to Development Studio and Deployment Tool are read from the config by the installer.

Updates are delivered by a Claude Code hook, a script that runs when a session starts. It detects a new version, offers to pull the changes, updates dependencies, and verifies that the tools launch. Local settings are preserved, conflicts with custom changes are shown, and rolling back to a previous version is possible.

First run on a clean machine

The hook was not the first step. First the author set up a clean virtual machine, cloned the repository, and launched the agent. The errors started before the first useful task: dependencies were missing that he had installed on his own machine long ago and forgotten. One tool needed Python, another Node.js, a third lacked libraries. At first the agent helped with the setup itself, but sometimes preparation consumed the entire five-hour Claude usage window.

Now a new participant runs the installer: it checks the environment, installs dependencies, and configures the MCP servers. An interrupted installation can be rerun, and completed steps are skipped. If several people hit the same error, its fix goes into the installer. It takes into account the operating system, the DBMS, and the platform version. For example, the platform works with both PostgreSQL and MSSQL, but the author's psql-based utilities are unavailable with MSSQL, and the installer warns about this. Tools are verified from within Claude Code, because working in a terminal and being accessible to the agent are different things. And since Bash ran unreliably under antivirus on Windows, the script entry points were moved to Node.js.

To ensure that teams' changes don't break the agent's behavior, evals are run: test scenarios with predefined success criteria. Say someone modified the base skill for their project, and it stopped performing a mandatory check before a task was completed. Such a scenario will expose the regression. The catch is that evals only cover the cases they describe.

What to do in practice

A harness for one person and a harness for a team differ not in the set of tools but in how much depends on the author. If you plan to share an agent harness, test the installation on a clean virtual machine, remove absolute paths, separate shared, project, and personal data, set an attempt limit on fix loops, and write evals for the key skills. A caveat: this is the experience of one engineer at one company, and solutions for other platforms will have to be worked out anew.

Story

Key Questions Answered

What exactly was being handed over?

The author receives requirements from clients and needs to quickly show a prototype of how to implement them on the platform. The Directum RX development environment (Development Studio) has a JSON-RPC API, which is connected to Claude Code via MCP. The agent creates entities, properties, and method stubs through the same mechanism a developer uses in the studio window, then writes the business logic in C#.

What are the key details on Shared server or local installation?

The author considered two options. The first is a shared service with sandboxes (such as GoClaw or OpenHands), where the user gets a preconfigured environment with the required product version. For a single team with one environment, this is convenient. But the harness had to serve several teams with different products, platform versions, and solution sets. One environment for everyone can't be assembled, and supporting multiple variants would require a separate team.

What are the key details on First run on a clean machine?

The hook was not the first step. First the author set up a clean virtual machine, cloned the repository, and launched the agent. The errors started before the first useful task: dependencies were missing that he had installed on his own machine long ago and forgotten. One tool needed Python, another Node.js, a third lacked libraries. At first the agent helped with the setup itself, but sometimes preparation consumed the entire five-hour Claude usage window.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…