Architecture writing for the people who have to approve, deploy, and eventually own these systems — security architects, platform leads, and the engineers who inherit whatever ships.
All 27 articles
What was built, and what each layer defends against. Implementation documentation rather than argument.
A control plane that holds no customer data, an execution plane inside your VPC, and a hypervisor boundary between agent-generated code and everything else. What each layer defends against, and what the shape costs.
Read articlePolicy the model cannot read, a tool roster compiled per turn, and a rendered blast-radius diff pinned to a version token. Three checks in the execution path, and the throughput ceiling they buy.
Read articleThe model writes the answer; the backend decides what it may point at. A grounding critic, marker resolution against the execution record, and why rendering a citation is a security boundary.
Read articleOne per environment we have actually deployed into, including the failure modes that only appear against a real instance.
Schema injection collapses somewhere past a few hundred tables. A narrow probing tool set, retrieval-driven discovery, a pointer pattern for bulk results — and the three failure modes that only appear once a real catalog is behind it.
Read articleTools split by consequence rather than by endpoint, a credential the sandbox never sees, and the two operations — pipeline triggers and stale approvals — that break a naive integration.
Read articleRole-scoped visibility breaks more dashboards than any other single cause. What the warehouses share, what they do not, and why the profiling step decides your deployment topology.
Read articleThe alternatives you are evaluating against, including where each of them is the better choice.
An orchestration library is not an alternative to a governed system — it is a component inside one. Twelve decisions the library leaves to you, and when owning them is the right call.
Read articleFour architectural decisions you inherit with a managed agent runtime, the one written commitment that rules it out, and why the switching cost is not where people look for it.
Read articleAn assistant and a governed execution substrate are not competing for the same budget line. The boundary sits exactly where an action stops being reversible — and inheriting a permission set is not the same as intersecting one.
Read articleMany organisations should build this, and the ones who should not are identifiable by one property of the work. The demo is the first fifth; the substrate is the rest, and it cannot be retrofitted onto a prototype that already works.
Read articleThey inherit governance rather than asserting it, which is why they win for analytical work inside one platform. They also stop where the warehouse stops, and most enterprise workflows do not.
Read articleCoordination frameworks solve delegation well. Delegation is also how an injected instruction escalates privilege, and agent count is a design cost rather than a capability.
Read articleThe arguments underneath the architecture.
The pilot worked. Eighteen months later it still is not in production. Five architectural questions decide that outcome, and all five are settled before the first line of code.
Read articleThe defence and the attack share a channel, so no instruction closes the class. Four named attacks, and the architectural layers that make a successful one harmless.
Read articleResearch and volume trade against each other, and every team eventually picks the losing side. The way out is to automate the research and leave the send alone.
Read articleGuessed addresses cost more than the campaign that produced them. Three escalating verification tiers, and the design decision that matters most when none of them can answer.
Read articleDeployment topology is not an infrastructure detail you settle later. It determines which data the system may touch, and it is the hardest decision to reverse.
Read articleMost human-in-the-loop is a confirmation dialog. Three properties separate a real approval gate from a checkbox that manufactures consent.
Read articleEvery demo works against twelve tables. Real catalogs have tens of thousands, and the standard approach collapses on cost, latency, and accuracy at the same time.
Read articleThe traceability matrix is accurate the day it is written and wrong within a sprint. Generating the link at authoring time is what makes it survive.
Read articleMost organisations can report line coverage to two decimals and cannot say which implemented behaviour nobody has written a test for. Those are different numbers.
Read articleThe assertions in a decade-old suite are usually fine. What is missing is everything around them — and retrofitting that is tractable in a way rewriting is not.
Read articleAdding analysts to a forty-item queue buys a quarter of relief. The queue is long because the same dashboard is being hand-built for the eleventh time.
Read articleThe question is not which tool is better. It is which numbers the board sees when four departments each maintain their own version of revenue.
Read articleFinance rejects AI business cases for predictable reasons. Understated loaded cost, unfalsifiable benefits, and an automation rate nobody believes.
Read articleA dashboard nobody opens is not a design failure. It is a specification failure — it was built from the data that existed rather than the decision it serves.
Read articleThe build-versus-buy framing hides the option most enterprises actually need, and the three constraints that decide it have nothing to do with engineering capacity.
Read articleBring the constraint that has been blocking you to a technical briefing. Forty-five minutes with the engineers who build the systems.
Book a Technical Briefing