Practical Agent Engineering for South Africa
An agent that works in a demonstration still needs two kinds of engineering before it can be trusted with a service. Someone has to decide which model handles each piece of work, and someone has to own the loop that calls the model and runs its tools. This series covers both for South African executives, architects and engineers. Part 1 treats model routing as a session-aware control decision: hard policy gates before any router, routing at the user-turn boundary, the cost of losing the prompt cache, and measurement by cost per accepted outcome. Part 2 sets out the runtime controls around the loop: budgets, typed stop reasons, hooks, the difference between cancelling a run and interrupting it for approval, checkpoints, redaction and the trace. It follows Gamtoos Valley Municipality, a fictional Eastern Cape municipality, through its citizen help service and its burst-pipe dispatch. Both parts draw on a small public-safe lab with synthetic data and simulated models, so its numbers show how to measure, not how any real model performs. Research claims were checked on 5 November 2026.
2 articles
- Part 1Stop sending every agent task to the most expensive modelThe cheapest model is not always cheaper, and the strongest model is not always needed. Part 1 of Practical Agent Engineering for South Africa treats model routing as a sequential control decision: hard gates first, session state in view, cache effects counted and cost measured per accepted outcome.9 min read
- Part 2You do not control the agent until you control the loopAn agent is not a model plus a list of tools. It is the loop that decides when the model runs, which tool may execute, when a person must approve and why the run stops. Part 2 of Practical Agent Engineering for South Africa sets out the runtime controls that make that loop operable.10 min read
Working through any of this in your own organisation?
These articles come out of delivery work rather than theory. If the problems in this series look familiar, a short scoping conversation is usually the quickest way to work out what is worth doing first.
Talk to CloudNala