Explore Explore Explore Explore

[jsirex]

eca is wonderful tool and worked really good. But with recent versions I noticed strange behavior. Which is simple wastes tokens and time. eca is obsessed by explore.

In plan mode I asked something close to:

there is all logic in src/proj/upload.py
look at this and also look src/proj/cli

make the same way `proj download --flag 1 --flag 2`

eca spawns subagent, it starts full explore and basically read all files, creates detailed description of each file,
For a project with 6 files 40K tokens generated wasted. Agent finishes and looks like all that data goes away.

eca thinks a little, decides "i will inspect file1 file2" and stand another explorer. Why? So files read again in another subagent.
and this continues 4+ times. Usually, I stop it and just ask to implement.

Sometimes, even if I say while cursor in file1 "in the file1 there method, update it", eca stars to explore whole project, dump files, summarize, etc...

90% of my prompts start from "Do not explore" or "Do not explore whole project, but look only into file1".

It feels, like I should never use plan and ask anything directly in code mode.
What am I doing wrong?

//I use gemini 3.5 flash.

[jsirex]

More cases here: even in code mode I explicitly asked ready this 2 files and implement:

1. Subagent started with explorer. file read
2. Subagent started again with explorer and file read again
3. Now work started and file read again..
   tokens wasted

[ericdallo]

@jsirex that's most likely a model issue, each model can behave differently so we have no control whether the model will decide to call a agent or not, the best you can do is tune the prompts of each agents, including plan agent if you want, not sure theres much we can do to improve that on the default prompt c/c @zikajk

[zikajk]

The Plan agent is really designed to perform extensive exploration and research—ideal for brainstorming before larger implementations. Though to be fair, newer models need it less and less, so I wouldn't use it if you're trying to minimize token usage.

It’d be great to see if tweaking the system prompt for the explorer sub-agent improves this behavior. If you experiment with that and find any solid improvements, please share!

Regarding the main issue: it could be interesting if the main agent could re-query the Explorer instead of re-doing the work. But there are a few blockers:

• Context tracking: How can the main model know what the Explorer already knows to reliably decide whether to re-use it?

• Cost effectiveness: Cache hits make it cheap, but if the cache is invalidated, resuming the Explorer gets more expensive than just having the main agent (or a new Explorer) re-read the file.

I could see this "resume" approach working as a user-promp-triggered action, but making it automatic seems problematic.

Also, I've seen Gemini 3.5 Flash do a ton of redundant reading—it really excels at over-reading files. That said, something in ECA might be contributing to it.