We have been trying to get Claude Code subagents to work well on a big production codebase. So far we haven’t had much luck.
Most of the time Claude doesn’t bother using a subagent even when the prompt is (IMO) pointing in that direction. When forced to, things get done much slower, and the final output is pretty similar in subjective quality.
It’s possible we are doing it wrong, but maybe we’re also observing the phenomenon you saw - since Claude wrote all our subagent config, it might just be saying “I already know all that stuff without needing to fork into separate agents”.
That makes sense; Claude Code recently added the ability to tag subagents with @ to force them to execute, so I think they are running into similar limitations internally.
It's a framework I started to define/design to use on an initial application I was developing purely through coding agents. This was before I really knew what it took to have a reasonable experience using nothing but agents to code the web app. I've paused work on the framework to dive 100% into solo dev web app coding so I've learned a lot that will eventually make it back into the framework. At the same time, things like Spec Kit and Agent OS have gained popularity to push the spec-driven development approach. And Amazon just released Kiro IDE as well.
Fascinating research!
We have been trying to get Claude Code subagents to work well on a big production codebase. So far we haven’t had much luck.
Most of the time Claude doesn’t bother using a subagent even when the prompt is (IMO) pointing in that direction. When forced to, things get done much slower, and the final output is pretty similar in subjective quality.
It’s possible we are doing it wrong, but maybe we’re also observing the phenomenon you saw - since Claude wrote all our subagent config, it might just be saying “I already know all that stuff without needing to fork into separate agents”.
That makes sense; Claude Code recently added the ability to tag subagents with @ to force them to execute, so I think they are running into similar limitations internally.
I feel every bit of this. I’ve attempted, poorly, to help agents maintain intent and context. Spoiler alert: it’s TOUGH. https://github.com/chavezabelino/aegis-framework
Can you give more context on that repo? You've been using it?
It's a framework I started to define/design to use on an initial application I was developing purely through coding agents. This was before I really knew what it took to have a reasonable experience using nothing but agents to code the web app. I've paused work on the framework to dive 100% into solo dev web app coding so I've learned a lot that will eventually make it back into the framework. At the same time, things like Spec Kit and Agent OS have gained popularity to push the spec-driven development approach. And Amazon just released Kiro IDE as well.
TL;DR plan, review, design, review, code, review.
"Engineering is driven by better algorithms, cleaner data, and more efficient memory usage: training-time self-improvement."
Can you expand on cleaner data? Esp. given dead internet-esque forces
Great research .. please see my article where I referenced your work https://medium.com/@anandglider/current-ai-systems-are-trapped-in-an-eternal-present-424cd5d48682