Honest answers
to hard questions.

A running AI agent harness, not a report or a deck. At the end of an engagement you have a system — written behavioral specifications, skills that encode your operating rules, and agents that execute them inside your workflows.

The deliverable is defined up front. You read the spec before the harness is built. If the spec is wrong, you say so before we write a line of code.

An invariant is a rule your operation has always followed — implicitly, repeatedly, reliably — that was never written down. A meta-rule is a higher-order version: the rule that explains a cluster of smaller rules.

Concrete example: a media-buying team always re-prices a placement if the reach estimate drops by more than 15% within 48 hours of a campaign start. Nobody documented that. It lives in their heads. When a new buyer joins, it takes months to learn it. That is an invariant. We find it, name it, and encode it so an agent can apply it without prompting.

The engagement is the process of finding those rules in your operation's actual history — not in interviews, not in the org chart — and writing them down precisely enough to run.

You own the specifications and the harness configuration built for your operation. The deliverable is defined at the start and agreed in writing; there is no ambiguity about what done looks like.

If a criterion in the specification is not met, we keep working until it is. The engagement does not end on a timer — it ends when the defined deliverable exists and runs.

An open-source release is in preparation. The harness is built in the open, and the methodology is public; a packaged release for general use is not yet published.

Engagements are built on this same harness — not a different, proprietary system we keep hidden. We would not sell a system we would not run.

No. The harness runs on your existing infrastructure. The specifications are readable text; a non-technical operator can see what each rule does and why.

Maintenance is a function of how often your underlying rules change. Stable operations with stable rules require minimal upkeep. For operations where rules evolve frequently, we scope a standing review cadence into the engagement.

Four to fourteen weeks, in three phases: Discovery, Specification, Harness.

Discovery (typically two weeks) is where we read the operation's history and surface the invariants. Specification (typically two to four weeks) is where we write the behavioral contracts — you review and approve before anything is built. Harness (the remaining weeks) is where the agents are built, connected to your workflows, and validated against the criteria in the spec.

The range exists because operations differ in complexity. The scope is fixed before the engagement starts; we do not add phases or deliverables mid-engagement without agreement.

An AI-fit diagnostic starts at $5,000 (¥30,000). The full workflow-to-system engagement is fixed-scope and quoted after a short scoping call — the scope (the number of workflows, the depth of the specification, the complexity of the harness) determines the cost.

The engagement deliverable is defined up front. You know what you are paying for before you commit, and the price does not change unless the scope does.

A fit: computer-based SMBs of roughly 15–150 people with no in-house AI team, fixed processes and repeatable decisions worth encoding, and a workflow painful enough to be worth the engagement.

Not a fit: teams that want prompt training or tool tutorials, operations with no process records to read, work that is entirely one-off or creative-from-scratch, and organizations without a decision owner or unwilling to let rules be written down.

Sectors where engagements run well: advertising and production, game development, cross-border e-commerce, logistics, B2B services, and ops-heavy teams.

You can. Building a harness internally requires someone who understands your operation deeply, can write precise behavioral specifications, and has the engineering skill to implement and validate them. If that person exists in your organization, the internal build is a real option.

Most operations that reach out do not have that combination. The discovery methodology — reading operational history to surface invariants — is not something a generalist AI engineer or a new hire picks up in a sprint. It is the core of what we do, and it takes the full engagement to do it correctly.

A big consultancy delivers a report. We deliver a system that runs. The diagnosis and the implementation are the same engagement; there is no "phase 2 handoff to an integrator."

An AI vendor sells a tool. We encode your rules into that tool. The vendor's tool is mechanism; the specification is the intelligence. Buying the tool without the specification leaves the rules still living in heads.

A dev shop builds what you specify. We write the specification. Most operations cannot yet specify their own behavioral rules in the precise form an agent can execute — that is the discovery and specification work we do first.

Engagements are conducted under a confidentiality agreement. We read your operational history — decision records, process logs, internal communications relevant to the workflow — under that agreement, and we do not retain or share it beyond the engagement.

Case studies on this site are anonymized or illustrative. We do not name clients without permission.