AI agent keeps getting worse as we add more capabilities

[Tatarum]

We started with a fairly simple AI agent that handled one specific workflow, and it worked surprisingly well. The problem is that over time we've kept adding new instructions, tools, fallback rules, edge cases, and special handling requests from different teams. Now the agent technically does a lot more than before, but the quality feels inconsistent. Sometimes it follows the wrong workflow, sometimes it ignores important context, and sometimes it overuses tools when a simple answer would have been enough. We can tell the behavior has changed, but it's becoming difficult to figure out which modifications actually improved the agent and which ones introduced new problems. How are teams evaluating and optimizing complex AI agents once they start accumulating lots of instructions and capabilities?

[Bening]

The consistency issue is almost always a signal that the instruction set has grown faster than the evaluation framework around it. Every team adds their piece thinking it's a small tweak, but the model is now trying to reconcile twenty different priorities at once and picking whichever one feels most salient in the moment. The fix isn't pruning instructions, it's building proper regression testing so you actually know what changed between versions. Run the same set of representative inputs before and after every modification and compare outputs systematically rather than spot-checking a handful of cases and calling it good.

[Galileo]

It's worth noting that a lot of teams hit this exact wall around the same point in the agent lifecycle, when it goes from a focused tool to something that's trying to serve too many masters at once. Getting the evaluation layer right is what separates agents that scale from ones that just accumulate technical debt. You can do your agent optimization here . The tooling there is built specifically for this kind of complexity, where you need to track behavior across versions and isolate which changes are actually moving the needle. Makes the whole process a lot more systematic than trying to eyeball outputs and guess what went wrong.