Building Governed AI Agents
Plenty of organisations are moving from chatbots to agents. The difficult part was never getting the agent to answer. It is making the thing reliable enough to trust, correctable when it is wrong, measurable when it changes, and affordable when it is used properly rather than demonstrated. This series works through both halves: how to design an agent people will actually use, and what it takes to run one without the bill arriving as a surprise.
17 articles
The series has two halves that are usually owned by different people. Parts 1 to 8 are about agent design: what the system does and how it gets better. Parts 9 to 15 are about what it costs to run and how that becomes visible. Parts 16 and 17 argue that these are the same readiness question, and put the whole thing in order.
Designing an agent people will use
Defaults instead of a text box, the tools people already work in, and a domain narrow enough for the system to learn what good looks like.
- Part 1Smart defaults beat infinite customisation in AI agentsYour users are worse prompt engineers than you are, and that is not a criticism. It is a design constraint, and the products that respect it get used.7 min read
- Part 2Good AI agents meet users where they already workMost AI adoption failures are not model failures. They are the entirely predictable result of asking people to change how they work before showing them why.7 min read
- Part 3Why strong AI agents start vertical before going horizontalA generic business assistant can help with anything and is trusted with nothing. Depth in one domain is what teaches a system what a good answer looks like.7 min read
What makes the behaviour reliable
Curated examples, workflow specs written as instructions, and a review loop that produces data rather than a thumbs-down.
- Part 4Examples are the real fuel of AI agent qualityTeams spend months tuning prompts and weeks arguing about models. The thing that most reliably improves an agent is a maintained library of what good and bad look like.7 min read
- Part 5AI agents need workflow specs, not vibesNatural language has quietly become a specification language. Most teams are still using it as a wish, which is why their agents behave differently every run.7 min read
- Part 6Human-in-the-loop is not a bottleneck, it is how AI agents improveHuman review is usually described as the temporary scaffolding you remove once the agent is good. In serious workflows it is the mechanism that makes the agent good.7 min read
The reliability layer
Memory as a governed store with rules and expiry, and evals wired to a release gate rather than read alongside one.
- Part 7AI agent memory must be designed, not accidentally accumulatedMost agent memory is whatever happened to be said, kept forever, owned by nobody. Five words fix the design: remember, ignore, expire, correct, audit.7 min read
- Part 8Continuous evals are tests for AI systemsAgents regress quietly and sound just as confident afterwards. Without a suite of known cases, you cannot tell improvement from a change of tone.7 min read
Where the money actually goes
Why the model price card is usually the wrong place to look, and why a healthy-looking cache metric can sit on top of a serious problem.
- Part 9AI coding costs are not just model pricesThe bill arrives, everyone looks at the model price card, and the wrong lever gets pulled. The cost usually lives in how much context each turn carries.7 min read
- Part 10Why a 98% cache hit rate can still hide an expensive AI workflowCache hit rate measures how much repeated context was discounted. It says nothing about whether you should have been carrying that context in the first place.6 min read
Instrumentation and habits
The per-turn view that turns "this was expensive" into "this turn was expensive", and the two habits that follow from it.
- Part 11Before you optimise AI coding costs, instrument every turnA session total tells you that something was expensive. Only a per-turn view tells you which turn made it expensive, and therefore what to change.7 min read
- Part 12Session hygiene is the cheapest AI coding cost controlOne task, one session, one short handoff. It costs nothing, requires no platform change, and usually beats every other cost intervention available.6 min read
- Part 13Subagents are a cost-control pattern, not just an AI architecture patternSubagents are usually explained as a way to decompose reasoning. The more immediate benefit is that they keep large reads out of the context you will pay for repeatedly.6 min read
Running it as an operating model
The weekly numbers, the three audiences who need different views of the same spend, and what a governance forum should actually read.
- Part 14Four AI engineering metrics teams should measure every weekCost, quality and acceptance have to be read together and weekly. Any one of them alone will let you declare victory while the other two get worse.6 min read
- Part 15AI cost observability needs three views: developer, team and financeThree audiences ask three different questions about the same spend. Building one dashboard and expecting it to serve all of them is how the numbers lose credibility.6 min read
Bringing it together
Why quality and cost are the same readiness question, and the fourteen steps from one workflow to a system a business can depend on.
- Part 16Strong AI agents need both evals and FinOpsAn agent is not production-ready because it gives good answers. Four questions have to be answerable at once, and three out of four is a specific, predictable failure.6 min read
- Part 17The CloudNala production agent playbookFourteen steps from one workflow to a system a business can depend on, in the order they pay off. Only the first one is needed to start.8 min read
One workflow, done properly, is the whole starting point
Most agent projects that stall did not pick the wrong model. They picked too broad a scope, wrote no spec, collected no examples and had no way to tell whether last week’s change helped. A short scoping pass on a single workflow usually settles what is worth building and what it will cost to run.
Book an AI Readiness Workshop