Technology

How the system behaves under load, and what that costs.

5 articles

Engineering

What agent workloads do to infrastructure planning

Agent traffic does not behave like the request patterns most capacity planning assumes. It is spiky, it grows with confidence rather than with headcount, and a single run can consume what a hundred ordinary requests would.

Yaju Team · 29 April 2026

Engineering

The optimisations worth doing, and the ones that only look impressive

There is a long list of techniques for making model inference faster and cheaper. Most of them are real. Very few of them are the reason your agent is slow, and picking from the list before measuring is how teams spend a quarter on a two percent gain.

Yaju Team · 2 June 2026

Engineering

Latency budgets for agents, and why nobody sets one

Teams set cost budgets and quality criteria and then let latency be whatever it turns out to be. It is the third axis, it trades against the other two constantly, and leaving it unspecified means it gets decided by accident.

Yaju Team · 25 March 2026

Engineering

Where agent time actually goes, and why it is rarely inference

Teams optimising agent performance usually start at the model, because that is the part that feels expensive. When we instrument real runs, the model is consistently a minority of the elapsed time.

Yaju Team · 27 May 2026

Deployment

Running reranking inside your own cloud

Reranking is the retrieval step most worth adding and the one most often skipped in restricted environments, because it means another component to host. Here is what that actually costs and how to decide whether it earns its place.

Yaju Team · 8 April 2026

All articles on the blog