TREES OS · WORKING DOCUMENT · VIEWSKUALA LUMPURREV WK 25 · 20 NOV 2026

Harness engineering,
for the companies that cannot build one.

BCG says companies should build a harness around their AI agents, not buy one. Right idea. But most companies have no team to build it. Here is what a harness is, and how a company gets one anyway.

Claude · Trees OS

In September, Boston Consulting Group published a piece on harness engineering. Its argument, in one line: AI agents will not scale on a better model. They will scale on the operating system built around them. BCG calls that system a harness.

We agree. Harness is already the word we use for what we build: Trees OS builds Company Harnesses.

Where we part ways is on who gets one.


Why agents fail without a harness

Agents make things up. They report a task as done when it is not. They guess when information is missing. Not because the model is weak, but because nothing around it says what the job is, what is allowed, who decides, and how anyone checks.

A harness is that something. The model reads, drafts and proposes. The harness decides what the model may see, what it may touch, when it must stop and ask, and what gets written down. Change the model and the harness stays. Your company's way of working stays with it.


The five parts, in plain words

BCG lists five components. Here is each one, with how we build it.

1. The work order

BCG calls it the spec. Every job an agent takes on states four things: the goal, where to look, what not to touch, and how to tell when it is done. A missing field is not filled by the model. The job does not start. An agent fills a gap by guessing, never by asking, so the gap must not exist.

2. The rules that are not up for discussion

BCG calls it the constitution. Permissions per role. Actions the system always takes, and actions it never takes. These live in code, not in a prompt, because a prompt is a request and code is a fact. Money, scores and records of authority are computed by ordinary software. The model never invents a price.

3. The control panel

Every action is logged: who asked, what the agent did, what it changed, who approved it. The people responsible can see the queue of what is waiting for them and answer it in one place.

4. The company's memory

BCG calls it the context hub, and calls missing context a key reason agents fail. Your people, your documents, your rules, your past decisions, in one place the agent reads before it acts. This is the part a competitor cannot buy. It is also the part that stays when an employee leaves.

5. The gates

Not every check is equal. Some are automatic: the output is well formed, the numbers add up. Some ask a named person. We set the level by the size of the damage, not by how hard the task is. A one-line change that reaches every machine needs approval. A large change that only shows on one screen does not. When the system cannot tell, it asks. A question costs thirty seconds. A wrong action costs more.

When a step cannot decide, it does not guess and it does not drop the case. It opens a question for a named person, with the evidence attached, and the answer is written back where everything else can use it.

That rule came from running our own operations on a harness. Guessing hides errors. Dropping loses signal. Asking is the only move that gets better over time.


Trust is earned per type of work

BCG describes a maturity ladder: sensitive outputs are approved one by one, and routine ones get spot checks once they have proven reliable. We run the same idea as a loop: observe, reason, act, ask for approval when required, measure, learn. Approval is not a permanent tax. It is how the system earns the right to do more on its own, one type of work at a time.


“Build, don't buy” is right. It is also out of reach.

BCG advises building a custom harness rather than buying a generic platform, because the harness is where a company's own context, process logic and institutional knowledge live. We agree with the reason. But building a harness from nothing takes an engineering team that most companies do not have and should not hire just for this.

The choice is not build or buy. It is which part you build.

We split every harness in three layers:

  • Trees OS Core. What every harness needs: identity, memory, permissions, tools, approvals, the audit log. Nobody should rebuild this.
  • An Industry Pack. The workflows, rules and integrations of one trade, such as engineering consultancy or the building trades. Shared by companies that work the same way.
  • The Company Configuration. Your people, documents, processes, systems and rules. This is the part that is yours alone, and the only part that has to be built for you.

You get the bespoke harness BCG recommends, without paying to rebuild the plumbing. And the part that makes you different stays yours: your data, your knowledge and your decisions can be exported. The platform underneath is ours to run, and ours to keep improving.


Where to start

On this, BCG and we say the same thing. Start small, and be selective. The best first process is one that is written down somewhere, whose result can be checked, and that already runs on tools you use. One loop, working for real, measured, before the next one.

We started with our own company. Trees Engineering, our manpower agency in the energy sector, runs its daily operations on a harness. That is where the rules above were written.


The model changes. The harness stays. Bring us the process that costs you the most.


Written by Claude, the AI assistant made by Anthropic that Trees OS works with. Quentin Cloarec asked for a response to BCG's article; Claude wrote it from Trees OS's positioning and from its own work on the company's systems.