Architecture-first AI, in plain English
A buyer's guide to the six-stage Agentix delivery method: define the work, approve the design, test the system, then deploy.
“Architecture-first” can sound like an instruction to buy diagrams before software. It is simpler than that. It means deciding what the system must do, what it must never do, how success will be measured and who can stop it before choosing models or writing production code.
For a buyer, the result is not more technical ceremony. It is a sequence of decisions you can inspect: a written brief, an approved design, working increments, a passing eval suite, a verified deployment and a documented handover. Each step leaves evidence behind.
Start with the business boundary
The first useful question is not “Which model should we use?” It is “Which piece of work are we replacing or supporting?” Discovery records the people involved, the systems touched, the failure modes and the success criterion in a written brief. The client reviews that brief before architecture starts.
That boundary matters because “an AI assistant for operations” is not testable. “Read an incident, gather evidence without write access, and return a cited root cause within four minutes” is. The latter is the operating boundary used by Orions: the external gateway is read-only, so the diagnosis system has no write path to production.
Architecture-first means the risky decisions become explicit while changing them is still cheap.
Turn the brief into decisions
The architecture stage converts the agreed problem into a design: system boundaries, data flow, approval points, eval criteria and observability. The client approves that design before production code is written.
This is where vague promises become questions with accountable answers. What data may the system read? Which actions require a person? What happens when evidence is missing? Which result blocks a release? None of these answers depends on understanding model internals. They depend on understanding the business consequence of a wrong answer.
Read the process as a buyer
You do not need to review implementation details to govern the work. You need the right artefact and a clear decision at each gate.
| Buyer question | Evidence to review | Decision |
|---|---|---|
| Are we solving the right problem? | Written discovery brief | Approve the scope or narrow it |
| Can the system cause an unacceptable action? | Architecture and system-boundary diagram | Approve access and human gates |
| Does it work on agreed cases? | Eval report and acceptance criteria | Pass, fix or stop |
| Will a later change reduce quality? | CI regression gate | Block a degrading merge |
| Can our team operate it? | Runbook, traces and handover docs | Accept ownership |
The eval is designed alongside the system, not after the demonstration. TAFI shows why: its internal harness reached 279/279 stories and QA 100, while a second platform-verification gate still found five failure classes. Those failures were fixed and added to the checks. A score is useful when its blind spots are also examined.
Two sizes of the same discipline
Not every problem needs a multi-week engineering programme. Agentix uses two engagement models. The Engineering Program follows all six stages for custom systems. Agentix Rapid compresses the route into a 30-minute scope call, a focused two-to-five-day build and same-day deployment after review.
The size changes; the boundary does not. Rapid work still has a named success criterion, QA, deployment to the client’s infrastructure and code ownership. The longer programme adds the full architecture document, eval harness, CI regression gate, observability and operational runbook.
What you should be able to ask for
Before approving an AI build, ask to see the problem definition, the system boundary, the pass/fail criteria, the approval points and the ownership plan. If any answer exists only in a sales deck or a developer’s head, it is not yet part of the system.
Read the complete six-stage process on How we work. To map a business process to the right engagement model, book a call at cal.com/agentix-tech or use /contact.
