First of all, thanks for sharing your excitement about the Agent 3 launch. It means a lot to me, as it motivates us even more to keep pushing the boundaries of what’s possible with coding agents.
Over the past week, you also shared a lot of constructive feedback with us. We’ve been reading every single post, and already began implementing several changes and optimizations.
I wanted to take a moment to address some of the feedback that has been submitted:
Reliability / Autonomy
Users have commented that the Agent 3 is too autonomous and does more work than they want. To rectify this, we launched an Autonomy Selector with 4 levels: Low, Medium, High, and Max. You can change your autonomy to “Low” to achieve an experience closest to Agent v2. For projects that you originally created with Agent v1 or Agent v2, we recommend “Medium” autonomy. “High” autonomy is the best option for new projects, as the Agent will catch potential issues immediately, hence working more reliably on longer runs.
We identified cases where App Testing was running for too long – fixes have been implemented to drastically reduce this. We are actively improving the tool, especially around login/auth, and more changes will be coming soon.
Pricing
We are quickly introducing more transparency with progressive updates on how much is being spent as the Agent works through the task list
Although we have not raised the pricing with Agent, 3 users have noticed that the Agent often runs longer, which ends up costing more as a consequence. You can expect to spend around $10/hr while agent is running autonomously (unless you enabled the High-Power Model). Keep in mind that amount could vary quite a lot, depending on several factors, including how much time is spent on App Testing.
While the cost for each user message may be higher, Agent 3 is acting more proactively, catching and fixing more bugs, preventing technical debt to fester. With Agent 3, you will often end up spending less to implement the same level of functionality.
Billing
The usage page is always the best place for aggregate billing information.
“End of run summary” is another useful resource, as it shows in detail the charges per task.
We are considering adding a running usage meter, so you can see how much the Agent has spent for the current run.
In general, we’re finding that new users are having a better experience with the added features of Agent 3, enabling them to create more advanced applications. However, we noticed that some older projects built with Agent v1 and v2 were not handled correctly by our latest release. This issue is not caused by an oversight in backward compatibility, but rather by the new improvements and advanced capabilities of Agent 3. As our new agent is eager to address technical debt, we observed that it could get overwhelmed by a codebase with several quality issues. As such, we created the Autonomy Selector, so you can choose the trade-off between cost and technical debt that makes you most comfortable.
Again, I would like to sincerely thank you all for building with Agent 3.
I will continue to listen to your feedback, as it will play a key role in shaping the product for all our users!
I have only had a chance to play briefly with the autonomy levels, and not with any hard core coding.
But the testing I managed to do was for a complete codebase audit (originally built in v2) - so it was only reading/analysing files. I was in High autonomy, and it performed well, only costing a few dollars in total (approx 8 mins run time).
The vibe I got from the agent felt like agent 2. It seemed to understand me and I understood it. I know it sounds mad talking about the agent’s personality. But when you’ve used it a lot, you just get a feeling. The initial v3 was speaking Klingon But my current feeling is you have hit the mark with the updates
My main concerns
there is no way on God’s green earth I will use Max autonomy (for a new or existing app). I have lost a little trust in Replit over the past 10 days, and letting Agent run for upto 200 mins with no control or visibility over cost is a ridiculous concept. I just do not understand why anyone would do this. But I will monitor others’ feedback as they try it
I am not a fan of Plan and Build modes. I don’t see the point and find switching between the 2 modes clunky. I prefer to stay in Build mode and instruct it carefully to not make any changes until I confirm. I did a small bit of this with the v3 updates (in High mode) and it followed my instructions to the letter
Also, I do not appreciate being charged $0.12 simply for agent to say “can you switch me from plan mode to build mode so I can carry out the changes”!
Small UI niggle
please rename either “High” autonomy mode or “High power mode” - they are confusing
Naming idea
if you need a name for the current Agent version (to distinguish it form the initial v3 which we’d all rather forget about), then how about Agent v3-va (variable autonomy)?
Thanks for the honest and transparent update @pirroh . Glad @realfunnyeric has been able to assist and it’s been amazing watching the whole of the replit team’s responses to the challenges.
Came on this morning to see the changes, liked the concept. Still having major cost and productive issues however. This morning, i wanted GA4 installed on my app, FB pixel added to it and to fix a flash reload issue i have. I put it in plan mode and took a prompt from Chatgpt setting out these planned changes and asked replit to put a plan together. Thought for 1 min charged me $0.50 for the plan. The plan was including like 15 other things it didnt need. I asked to simplify the plan to just focus on my changes. It just gave me an updated checklist but didnt explain what i was doing. 1 min work $0.33 charge. Then i decided to try it on medium autonomy, no app preview and to just see what it did. It deleted 3 components, built a custom redirect (not needed) and worked for 5 minutes not fixing but further breaking even more issues and I was charged $2.11. I had to rollback completely as if i never started the conversation. Whole process took me 30 minutes of my time plus the time to write this. Trying again but this is the problem, last week it bandaid fixed issues charged me over $25 for the work when it was in its broken mode. Now with these “improvements” i’m breaking more of the already broken logic and having to pay for it. It is nearly getting to the price that i could hire a human dev to employ these changes for cheaper. These are not complicated requests or features and replit agent seems incapable of producing the changes and is wildly expensive. 7 minutes of work brings my replit hourly to $25/hr. tracking in the past when it was v2 and doing better i was running $6-8/hr and was able to code up to 4 apps across 4 browser tabs. Now i have to babysit and watch every step it makes and still doing continous rollbacks of which i get no refund for renders this still terribly unusable. If this is just me than i’ll try and improve my process and workflow, but I feel like I still have a completely unusable version of replit?
he was saying its still not working for him and hoping it maybe hadnt fully rolled out for him yet, but if the new settings are there its rolled out. I’m now 4 hours later still trying to fix my issue still above… Replit agent has died and i will be mourning his v2 death
So it means when they release the updates and or improvements, not for the whole system as once, but scheduled in different parts or per different user groups or parts of the world.
I appreciate the update, but my experience differs from the claims. I tried both making changes to an older codebase that I started with Agent 1, and also building a new app entirely with Agent 3. Both are extremely expensive (more than the stated 10/hr). The app built with Agent 3 did not work at all, and it was stuck trying to make simple fixes. It feels a lot less fun than it used to be when I am spending a good chunk of money and getting mixed results. I will keep using Replit because I am used to it, but started testing other tools like lovable that are a lot friendlier in terms of price (I was already able to build some simple apps for free in the last few days).
Be very careful thinking that Lovable is cheaper. I can absoutely guarantee that once their VC dollars run out, they too will start hiking the prices.
I am not saying that to put Lovable or any of the others down. They are all amazing.
But they are all following the same VC-backed startup patterns. Lower prices, subsidised by venture capital, to hook in a very large customer base. And then the dials are turned…
providing an update here, from what I can tell, Replits fees are becoming more reasonable per prompt again. I have updated my workflow to include chatGPT having it help me spin up detailed prompts, fact check them for best practices and then input them into replit plan mode ask it to validate the prompt i built against my codebase, then replit usually has a couple tweaks, to which i send that back to ChatGPT to validate and once they both agree and it sounds logically accurate to me, I have it build. This has reduced the cost and time to develop even though it sounds like more work up front, you don’t force replit to build something that is going to be riddled with bugs and issues that you have to fix.
I think the workflows are always changing cause now we have “fast” mode for quick bug fixes which is currently free for the moment and very good at quick short bug fixes or changes.
But overall, I would say Agent 3 finally has earned my trust back and isnt wildly inflating prices with no effect. I dont really want to met agent 4 anytime soon though…
I would like to (and they are) see replit integrate more services natively. The use of ai right into your app is very nice for personal applications and the connectors are great too. I also am really excited about the mobile app beta launching right now. I think Replit is in a good state at the moment.
But my current feeling is you have hit the mark with the updates 
