It seems there’s a significant difference in how Agent 3 and Agent 4 communicated context and intent to their subordinate agents to accomplish tasks, and how the primary agent utilized that context to assess the work performed by those subordinate agents.
I’m seeing major breaking changes introduced in my project with frightening regularity.
Large and intricate tasks that Agent 3 could independently accomplish within an hour or two now require 4-6 hours of constant assistance, leading us to question the viability of continuing with Replit unless a significant change is made.
My team has decided to revert all changes made since Agent 4’s launch and take a week off while Replit works on resolving their issues.
I personally average around $500/day in spending on Replit. With Agent 3 that was worth every penny. Agent 4 is just not usable beyond cute, simple apps.
1 Like
Great insight @mikeamancuso. I gave it a try yesterday and had the same experienced. I now have a case open and am waiting on support to help me sync changes made back into the main version. One task that has been spinning in limbo for 22 hours is blocking everything. Fortunately, I had minor, but significant updates I was trying to push back to main. This rollout was very rudderless and lacked sufficient documentation and guidance for users. Not good.
I have been having the agent put all tasks into a single “mega task” and running this prompt after it finishes to keep it on course:
Can you give me an after action report of what we did, why we did it, how we did it, and where it happened in the codebase?
Also, let’s do a quick retro. Did you observe any opportunities for us to improve our code base or discover any deficiencies? Are there any lose ends from our work? Pre-existing bugs? Any brittle code or patterns that would struggle with production data? Did we do anything that would make it difficult for future parallel agents to swarm? Where are we at in the backlog?
I follow that up with:
Help me incorporate all of the retro items you identified into our plan. Business value is equal for all items, so sequence work by dependencies and order to maximize parallel execution.
As a fun aside, I use the logitech mx creative console to keep my common Replit prompts organized and easy to evoke.
