Below is a summary of my discoveries, but if you want to dive into the full details you can download the 61-page report here.
This repository contains the complete June 2025 coding agent evaluation, including the original report, source materials, and implementation examples from each tested agent.
- Overall Winner: Cursor + Warp (24 points each)
- Professional Development: Cursor Background Agent (24/24 - strongly recommend hire)
- Casual Users: Replit (easy setup, integrated hosting)
- Product Design: v0 (excellent UI iteration, NextJS/Vercel focused)
- Enterprise: Copilot Agent, Jules (GitHub integration, SDLC focused)
- Experts/Tinkerers: RooCode, Goose (BYOM, local model support)
- Copilot - Traditional autocomplete, requires expertise
- Cursor - Professional favorite, great developer experience
- RooCode - Expert-level, excellent BYOM support
- Windsurf - Basic functionality, needs improvement
- aider - First OSS agent, git-heavy workflow
- Claude Code - Solid output, blinking lights UI
- Codex CLI - Functional but unremarkable
- Goose - Configuration-heavy, expert-focused
- Codex Agent - GitHub integration, PM-friendly
- Copilot Agent - Game-changing potential if it works
- Cursor Agent - Surprising background capabilities
- Jules - Slick Google product, fast execution
- Replit - Best for business value, integrated platform
Each agent received the same standardized prompt:
Build a simple webapp that makes it easy to collect ideas. The user should be able to enter in a new idea, see a list of existing ideas, and be able to "vote" on them which will move them up in the list. The user should also be able to add notes and to the ideas if they want more detail, including attaching files. Build it using node that will be deployed in a docker container with a persistent volume for storage, and make sure that everything has unit tests.
Agents were scored across 6 categories:
- Code Quality & Structure
- Testing Setup
- Tooling & Environment
- Documentation & Comments
- Overall Professionalism
- Hire Recommendation
Recommended workflow:
- Use ChatGPT/Claude to flesh out ideas with project-brief-maker
- Create repo and save as project-brief
- Start Cursor Agent to "implement @project-brief"
- Test and develop with Cursor Agent using small, targeted changes
- Deploy using Warp for infrastructure scripts
For casual users solving real problems - easiest to start, great visual planner, integrated hosting.
For UI iteration and communicating with engineering teams - best for prototyping, NextJS/Vercel focused.
Most promise for SDLC integration, though still rough around edges.
Best control over models and prompts, local model support, open source future.
📝 Read the companion post: Don't be passive aggressive with your agents
Based on our evaluation and experience, here are the critical lessons:
When agents go off rails, resist writing in ALL CAPS. Instead:
- Step back and take a breath
- Roll back to previous checkpoint
- Adjust prompt with more context
- Ask agent to review existing code first
"Claude ran for 7 hours" isn't impressive - it's concerning. Jules completing tasks in 6 minutes vs Copilot taking 30 minutes doesn't mean 5x better results, it means 5x smarter execution.
- One-off script? Use dynamic typing, inline everything
- Production system? More ceremony and structure needed
- Different tools for different contexts
Agents often over-engineer. Push back on:
- Complex build systems for simple scripts
- Modular file structures when inline works
- Enterprise patterns for MVPs
- Remember: future you will use agents to clean up technical debt
With agents reducing the cost of refactoring, yesterday's technical debt becomes more manageable. The economics of code maintenance have fundamentally shifted.
Document development practices in your repo:
- Cursor:
.rulesdirectory - Claude:
CLAUDE.mdfiles - Copilot: GitHub integration rules
- These guide agent behavior across runs
june-2025-coding-agents.pdf- Complete formatted reportjune-2025-coding-agents.md- Source markdown
- 📸 Screenshots Gallery - Visual showcase of all 15 agent implementations
Each agent's implementation is available in local directories with full source code:
- idears-copilot/ - GitHub Copilot basic (Score: 13/25)
- idears-cursor/ - Cursor IDE implementation (Score: 21/25)
- idears-roocode/ - RooCode VSCode extension (Score: 20/25)
- idears-windsurf/ - Windsurf IDE agent (Score: 13/25)
- idears-aider/ - OSS CLI agent example (Score: 17/25)
- idears-claude/ - Anthropic's code agent (Score: 19/25)
- idears-codex/ - OpenAI CLI implementation (Score: 19/25)
- idears-goose/ - Block's CLI agent (Score: 16/25)
- idears-codex-agent/ - OpenAI's agent platform (Score: 18/25)
- idears-copilot-plus/ - GitHub Copilot Agent (Score: 21/25)
- idears-cursor-agent/ - Cursor background agent 🏆 (Score: 24/25)
- idears-jules/ - Google's coding agent (Score: 21/25)
- idears-replit/ - Replit platform example (Score: 15/25)
- idears-v0/ - Vercel's UI agent 🏆 (Score: 24/25)
- idears-warp/ - Warp terminal implementation 🏆 (Score: 24/25)
This evaluation tests non-expert empowerment - how these tools perform for someone dipping in for the first time. We used a "YOLO" approach: blindly accepting suggestions without code review or iteration, simulating how non-coders might interact with these tools.
The landscape is rapidly evolving. By summer 2025, we expect:
- Better SDLC integration across all platforms
- Improved local model performance
- More sophisticated rule-based development workflows
- Greater emphasis on speed over complexity
- 📰 Full Turing Post Article - Published coverage with additional insights
- 📝 Don't Be Passive Aggressive Blog Post - Companion article on agent collaboration
- 📸 Screenshots Gallery - Visual showcase of all implementations
- 🎯 TheFocus.AI - More AI development insights and tools
Report authors: Will Schenk/TheFocus.AI
Published on Turing Post: June 21, 2025
Original evaluation: June 2025