Five principles. One firm.

The default model in our industry is to scale by hiring. A win means a bigger team. A new client means another bench. Margins come from utilization, not from the work. The work is, in practice, a vehicle for the headcount.

We chose the opposite. The firm scales by adding leverage — proprietary agents, reusable playbooks, a platform that compounds — and by adding senior engineers selectively, around that leverage. The test we apply is concrete: if a hire only improves utilization, we don't make it. If a hire makes the platform better for every future engagement, we do.

The reason isn't moral. It's structural. A firm built on headcount has to keep its people billable; that pressure determines what kind of work it can say no to, which is to say, not much. A firm built on leverage can afford to be selective about the work, because the work isn't the only thing the firm is. That selectivity is the entire reason a client comes to us instead of someone larger.

The visible side of this principle is what we don't do: no bench, no internal sales org, no "we'll bring in a senior" stage gate. The invisible side is what it forces us to do well: every engagement has to either solve a real problem or improve the platform. Usually both. If it does neither, we shouldn't have taken it.

AI is now fast enough that a competent agent can produce more code in an afternoon than a senior engineer can read in a week. The temptation is to keep up by letting the agent ship — to make the human a reviewer of last resort, or worse, an approver of a summary. We don't do that, and we don't think anyone serious should.

The principle is that the gate — the moment something becomes part of the production system — is owned by a named senior engineer who has, in fact, understood the change. Not skimmed it. Not been briefed by the agent on it. Understood it the way a senior engineer understands changes they intend to be on call for.

To make that practical at the speed AI now operates, we built ArchTrace: a deterministic gate that fact-checks every AI session against the architecture graph before the agent stops. That's not a substitute for the human; it's what makes the human's job feasible. The engineer reads the deltas that matter and trusts the gate on the ones that don't.

This rules out a class of work we'd otherwise be able to win: "just have the AI do it overnight." We say no to that. The engineering judgment is the product. The AI is what makes it affordable.

Almost every firm in our category overclaims about AI. Some do it cynically, because the market rewards the claim. Most do it accidentally, because the people writing the marketing aren't the people doing the work. The result is a buyer environment in which everything sounds the same and nothing is trustworthy.

We treat that as an opportunity, not a constraint. On every engagement we name, in writing, the places AI carries the work — analysis, drafting, regression testing, documentation — and the places it shouldn't go near the keyboard: judgment calls, trade-off decisions, production cutover, the conversation with the on-call team at 2am. The plan reflects that split. The bill reflects that split.

The discipline is concrete: when a sales conversation goes somewhere we can't deliver, we say so on the call, not in the SoW. When an experiment with a new agent doesn't pan out, we tell the client. When a model gets better and a manual step becomes automatable, we tell the client that too — even when the immediate effect is fewer billable hours.

Calibration is the moat. A buyer who has been burned by overclaim three times in a row recognizes the opposite when they see it. We'd rather be the firm that won the engagement by saying what AI can't do.

Most consulting work is consumed at the point of delivery. The client gets the system. The firm gets the invoice. Whatever the firm learned — the agent that turned out to be unusually good at reading SAP Z-code, the test pack that caught the bug nobody believed existed, the migration playbook that survived contact with reality — leaves with the engagement.

We design engagements to keep that learning. Every engagement is structured so that some piece of it becomes part of the platform: a reusable agent, a workflow, a connector, a test pack, a piece of a playbook. The client gets the deliverable. The firm gets the deliverable and the increment to the asset base.

The mechanism is simple and we publish it: a small percentage of every engagement is reserved for extraction — the engineer time it takes to make the in-flight learning reusable rather than one-off. The client benefits because the next engagement starts further down the road. The firm benefits because the platform gets a little stronger every time it ships.

Compounding is the entire game. Without it, an AI-leveraged firm is just a faster body shop, and a faster body shop is a worse business than a slower one. With it, the firm gets cheaper to operate over time without getting cheaper to buy. That's the asymmetry we're optimizing for.

The firm is engineered to take fewer engagements than a firm of its size could. That isn't a stage we're in on the way to bigger; it's the model. A book of work with five deep clients gives every engineer in the firm context that a book of fifty couldn't. Context is the second-largest cost in our work. Most of the rest is correcting decisions made without it.

What "selective" means in practice: we say no to engagements where AI leverage isn't the differentiator, where the client wants a vendor rather than a partner, or where the scope can be solved by adding people. None of those are bad engagements. They are bad engagements for us. A larger firm will do them better and cheaper.

What it means for clients we do take is that the engineers on the engagement know the system. Not the project. The system: how it was built, how it's used, where it's fragile, where the next change should go. That kind of knowledge takes time to acquire and we can only afford to acquire it because we're not also trying to do twenty other things.

Volume is a strategy. So is the absence of volume. We picked the second one. The work we get to do because of that choice is the reason any of the rest of this matters.