Once you put more than one agent in a pipeline, the hard part stops being “how do the agents collaborate?” and becomes “what crosses the boundary between them?” Multi-agent work fails when every sub-agent returns its own shape, verbosity, assumptions, and context residue. The scalable unit is not conversation. It is contract. New post: Orchestration Patterns for Multi-Agent Work - patrickelmore.com/posts/mult…

Is anyone in the enterprise engineering world doing anything similar to a traditional test suite to validate changes to their AI harness? I've been bouncing some ideas around and was wondering if anyone had solved any of the problems I am sure I will run into.

The model does not sort the context out. The composition does. Why most agent "bad output" is upstream of the model, and the three patterns that catch it: source routing, self-containment, authority. patrickelmore.com/posts/comp…

Bigger context windows don't fix what most people think they fix. Agents lose the thread late in long sessions because the context fills with debris, not because it runs out of room. Capacity expansion postpones the symptom; it doesn't address the cause. New post on context pollution: patrickelmore.com/posts/cont…

The agent-pause assumption: humans pause when the spec doesn't make sense. Agents don't. They pick the most plausible value and continue building. This is why single-pass planning fails silently with agents in the implementer seat: patrickelmore.com/posts/tier…