I have been having mixed results with Agent 3, I have been going back and forth with everyone here and it seems I’m one of the few still having issues getting Agent 3 to do anything with my project. I was having some success for like two days i switched between medium and high and it was doing a few things and I had accumulated about $200 or so in charges with very little progress but I accepted that as part of the learning experience on the new system which I am ok with. For about 4 days now I have been trying different prompts to try to address things I still need to work on, they all fail, they start working then don’t do anything, it hangs and gives no feedback, I end up refreshing and when i ask it, please try again, it acts like its “dislodged” whatever thought process it was stuck on then it strats to analyze and gets stuck again. I have been doing this for days, it did do a few simple tasks, but after that it got stuck again and won’t progress with any edits or changes. During this time I see my charges are still going up, it has charged me for these attempts, all failed meaning none have completed or even given me feedback at all of any changes to the code or even evaluation. I check my usage today and the last 4 days i was charged around $70 in charges for it doing nothing. It sucks because I want to keep trying but it seems like i’m burning money attempting and i’m stuck not being able to get any work done. I am easily $5000-$6000 into this project since i started and now i can’t even progress, I don’t mind paying if its working but its just not doing anything. I submitted a support request last week, have not heard back except for the bot responses which said its as escalated to a team member who are backlogged. I have still yet to receive any response, as of right now i’m just stuck with an unfinished project and the new replit system obviously can’t work with my old code. I know the project is huge, but agent 2 was able to browse easily and I have a good documentation system on modules and everything, something is locking up the process during its handoff to the architect and back to agent and I can’t see the internal mechanisms of what’s happening so I’m at the mercy of replit to figure it out. Any ideas guys?
come on man simply shift to claude directly and avoid replit wrapper -
just use claude code in replit shell and boom , and before going ahead make /agents and create a architecture who reviews all code just like replit and there are standard agents also by claude and call after each code written and provide context with a file like xyz.md to maintain and preserve replit environment config files and keep unlimited opus 4.1 and sonnet 4 unlimited even better , faster and 100x cheaper coz its unlimited with pro max 200$ plan and i myself make like 55-60 apps a month with 200$ thing , i spent 4-5k on replit but now its a big no coz they are basically a wrapper which is expensive silly toy to me and the speed is what the real issue is to me , Direct Claude Code Integration in Replit Shell
Skip Replit Agent entirely and implement Claude Code directly through the shell interface for maximum speed and cost efficiency.
Core Architecture Setup
1. Shell Integration
-
Install Claude Code CLI globally:
npm install -g @anthropic/claude-code-cli -
Export API keys through Replit Secrets
-
Initialize with
claude-code init --project-type=replit --integration=shell
2. Agent Directory Structure
Create /agents folder with specialized review agents:
-
code-reviewer.md - Architecture documentation for code review workflows
-
standard-agents.md - Claude’s production-ready subagents configuration
-
context-preservation.md - Environment config file maintenance strategy
3. Context Management System
Implement automatic context preservation using .md files that maintain:
-
Replit environment configurations
-
Project state and dependencies
-
Code review history and patterns
-
Agent interaction logs
Cost-Performance Advantage
Current Economics:
-
Your production: 55-60 apps/month at $200 (Claude Pro Max)
-
Previous Replit spend: $4-5K
-
Savings: 95%+ cost reduction
Speed Optimization:
-
Direct shell access eliminates GUI bottlenecks
-
Sub-2s code generation response times achievable
-
No Replit Agent wrapper delays
Implementation Strategy
Node.js Orchestrator Bridge:
javascript
// claude-orchestrator.js classReplitClaudeOrchestrator { constructor() { this.claude = newClaudeCode({ apiKey: process.env.ANTHROPIC_API_KEY, model: 'claude-3-5-sonnet-20241022', maxTokens: 8192 }); } }
Custom Shell Script:
-
Daemon mode on port 3001
-
Environment variable automation
-
Seamless file modification workflows
This approach transforms Replit from an expensive wrapper into a deployment platform while Claude Code handles the actual development intelligence, giving you enterprise-grade AI orchestration at consumer pricing.
I saw you post about this the other day, are there any issues you encountered working like this? I am not a huge fan of the architect anyway. I will consider this. Isn’t agent 2 just running claude opus so its like going back to that with the added speed benefit?
How long before access to Claude Code seems to be mysteriously blocked!? After all, it is directly undermining Replit’s main revenue stream!
you raise a fair point, this seems like something they will plug soon, we should probably stop talking about it 
One more thing, i’m assuming that the checkpoint system will be invalidated, correct?
even it has a rollback and stage all commit in replit git before you confirm any changes
replit is lowkey dead as the speed difference is insane and why not unlimited opus 4.1 and sonnet 4 with all features just like replit in vs code with extension or in replit shell with npx claude ? why would we pay for insance markup , poor performance and too slow , i tried many times but replit dont work any more or a moderate project , so i would say at the end they are a crappy wrapper with all money going in marketing rather than product development , probably this company gonna die anyways as its already lost trust
its like 200$ with claude pro max with 20x capacity which offers 2Billion token with a month and works 100x faster and accurate tho ,
and they cant block claude code , everyone will migrate simply to get google 350k credits for backend and hosting for free for next 2 years and will make full project with claude code in vs code terminal or maybe with extension and
steve you are an old man , i would say you know better than me , but at the end Replit is basically a VS Code and i dont think anyone needed a VS code that is expensive and slow and unreliable in many terms , so replit is basically a VS code marketed as no coding ai tool but its crap and probably replit gonna die in near future and in past 2 decades i have seen fall of such poor companies , builder.io is recent example
Replit vs claude code in shell , replit agent is searching for past 13 min and its set on low autonomy tho and i tried with max also , its trash compared to claude code in shell
Not quite sure if that’s an insult or not? 
I’d actually say we’re as old as the code we write, and I last wrote some 2 hours ago. So feeling pretty young - but also with enough experience to know when someone is just out to trash something without listening to others.
conversation.end();