Category
Engineering notes for the people who build and run agents.
7 articles
Engineering
The interesting work in most organisations happens in systems that were designed decades before agents existed, under assumptions that agents violate. Connecting the two is where deployments succeed or quietly go wrong.
Yaju Team · 6 May 2026
Engineering
There are two ways to build an agent and most teams pick one without noticing they chose. The choice determines how the agent fails, which is more useful to know in advance than how it succeeds.
Yaju Team · 1 April 2026
Engineering
Maintaining a fork is unglamorous, never finished, and quietly expensive. It also has exactly the properties that make a task suitable for an agent, which is rarer than the enthusiasm around agents suggests.
Yaju Team · 21 May 2026
Engineering
Most teams pick a chunk size in the first hour of a project and never revisit it. It then quietly determines what an agent can and cannot find for the rest of that system's life. Here is how we think about it, and how to tell when yours is wrong.
Yaju Team · 3 March 2026
Engineering
Choosing an embedding model feels like picking a component. It is closer to choosing a file format: everything downstream assumes it, and changing it later means rebuilding the index and revalidating everything that depended on the old one.
Yaju Team · 20 May 2026
Engineering
The diagram that explains the architecture, the screenshot in the incident report, the photographed whiteboard from the design review. All of it sits in systems an agent can reach and none of it is searchable, because the index only ever saw the words around it.
Yaju Team · 7 April 2026
Engineering
Most retrieval systems return ten passages and hand all ten to the model. Adding a step that reorders them and keeps the best three is usually the largest single accuracy improvement available, and it reduces cost at the same time.
Yaju Team · 30 March 2026
