I wanted to thank the replit team for fixing the issues with Agent 3, I have been actively using it for two weeks now and we are back in business. I have been able to control cost again and it seems to perform well. I attached my screenshot when i refactored my login system, 22 minutes of work, under $7.
2 Likes
Iād agree. Build mode @ Medium autonomy has been my sweet spot for both planning and building. The Agent has been very āobedientā to instructions and guardrails in my prompts. Agent has been great about planning-only tasks within Build mode which makes the interaction flow very well (without having to switch modes).
When required, the Build Agent references the āArchitectā (which sounds like a Plan mode agent). This has proven to be a reliable resource for auditing code and planning features and fixes.
Accuracy and cost have been great and the overall experience has encouraged me to push ahead a little harder.
2 Likes
Iām also homing in on Medium autonomy as the sweetspot for build mode.
But why wouldnāt you use High autonomy when you are doing stuff in planning mode? It canāt edit/break anything, so surely the highest autonomy possible makes sense?
I actually started my journey with V3 doing that. I then began experimenting with the idea of finding a single mode that required less āmanagementā ie switching. I do move to High Planning on occasion for new feature planning. However, from the mode descriptions in the UI, it appears that a major difference between Medium and High is the breadth of the code base youāre giving it access to. Iāve found that, while planning, I can prompt the Agent to take a wider view of the code in Build/Medium without having to switch to Build/High.
A question I did have: do the autonomy levels actually apply to Planning mode? The descriptions sound like build parameters not planning parameters. Welcome your thoughts
1 Like
I had assumed so, which is why I like plan mode to use high autonomy, so it can read and understand the entire codebase as it is planning a new feature.
But now youāve asked the question, I realise I have absolutely no idea 
It would be interesting to see how each of the Build and Planning autonomy levels handle the exact same task so we could clearly distinguish the impact (8 different views of the same tasks). Written descriptions can be misinterpreted and trial and error can be expensive.
Iāve found v3 so āobedientā to my instructions that I am finally feeling comfortable again using specific prompting to drive its behavior rather than relying on built in autonomy levels. During one build, it encountered a conflict we hadnāt discussed during our planning discussions. Just as it was instructed to do, it paused the build, presented me with the problem and various solutions and asked for my opinion on how to proceed. It did not go off and assumed it had the right answer and had permission to do what it wanted. (it was set to Medium)
2 Likes
Yes, big thanks to the team, Iām in production ahead of schedule even with the week and some change of being out of commission!
Now onto my complaining about managing environments on Replit 
In all seriousness though, it feels like Iām just one step ahead of the product direction and the releases come quick to resolve the walls I run into. Hope this continues!
3 Likes
Yes. This! 
I believe this kind of prompting is key.
Donāt tell the agent to disappear for 3 hours and return with a magically finished app. It will never happen and this is not how any real developer works.
Instead, ask it to āwork through our plan incrementally in phases, and pause at natural checkpoints to engage in a 2-way conversation with me to look over whatās done, fix things, add new secrets, run some DB schema SQL, test, discuss and address issues, etc, etc, etcā¦ā
2 Likes
Yes, there is a significant improvement and back to business. The costs are sensible now and the output is great. Thank you Replit team.