I feel this Reddit post is well rounded and raises a concern a few people are starting to worry about: that Replit is maybe heading too far down the point & click vibe coding route and removing too much autonomy from users, particularly the professional builders amongst us.
It would be good to get other people’s opinions, and as an ambassador I’d be happy to pass these along to Replit…
I have found that when you use plan mode in Replit, it does what you are describing. Agents come out of nowhere, a task list is made, you have questions to answer etc It becomes a mess and goes off an on tangent normally.
I now days only use “Build”, on “High” with “No app testing” with high level prompts going in from a parallel Claude chat that has the complete build planned. I never use “plan” anymore as it burns your money, plus adds frustration as pointed out in the original article with agents.
@Gipity-Steve I actually still see Replit as one of the best tools out there for building fast and experimenting with ideas. The Agent has been incredibly useful for me personally, especially when I give it clear direction and work with it step by step.
I think Replit is pushing hard toward making software creation more accessible, and honestly that’s exciting to watch in real time. No tool is perfect, but the speed, convenience, deployment flow, and overall ecosystem are still miles ahead of what many platforms offer today.
Since i saw Armjad’s excitement from agent 3 era advertising the autonomy behaviour i knew this was too ambitious for the time, even today… it is generally a bad idea to let the agents build autonomously for long periods they are going off rails too easily, so i believe that this was the right choice , usually 10-15 minutes is a good med long run to not screw a lot of things up and a safer approach.
The ambition though is becoming reality as newer models are coming out, i believe that this might be achieved once anthropic takes more gpus online through their newly acquired datacenters.
I totally welcome this safeguard, they are putting them to protect us from disasters
This is a great question and Reddit post.
I think the 40+ years of personal computing shows higher levels of autonomy is where we have always been headed. However, I agree with @Luram , Replit might be offering too much autonomy for what the tools can safely handle right now and too much autonomy for most vibe coders (including me).
I have wrestled to find a good rhythm with Agent 4. I tried the new planning/task management features but even the simplest task seem to get over complicated with Agent 4. I am shifting back to my approach with Agent 3: detailed written specifications that I work on with the Agent and validate with the Architect. The task plans created by Agent 4 were too general and left too much room for the AI to wander. In some cases it committed code without my approval. With my written plans I insist that the Agent specifically writes out the code changes it intends to make so there is no room to wander.
My execution approach is similar to @grayhill5 - build on Power with no app testing. A narrowly scoped plan, executed aggressively worked well with Agent 3 and I am hopeful I can get back to that rhythm.
I agree with issues in plan and build mode for replit. I have not used plan mode inside replit. Itsgood to plan product, flows, integrations, develop the plan into phases. All of these are done outside of replit. Then keep inputting phases and keep building the app. It worked better for me and keeps cost in control.
I asked the agent about making code changes in a fork of my main code base so I could review the changes prior to pushing them to the main code base. While this is now possible with the recent updates, the Agent offered a caveat with this approach:
The task agent runs autonomously — once it starts, I can’t course-correct it mid-flight. So the plan needs to be as complete as it currently is (which the architect just approved). Any clarification or scope tweak you want should happen before you hand it off.
This might explain how the autonomous builds go off the rails. Once it starts…it can’t stop. My standard routine is different: when I give the Agent the go-ahead to execute an agreed upon specification, I always tell it to pause and ask me questions if it encounters any issues that we haven’t previously discussed. About 20% of the time it will pause, we resolve the questions and it continues.
But if the autonomous ‘task agent’ has to run from start to finish without pausing, based solely on the plan in front of it, it is completely understandable how it will execute changes that might not be advisable. It is optimizing for task completion.