How to Evaluate TRAE's Release of 2.0 and the New SOLO Mode Experience?

How to Evaluate TRAE's Release of 2.0 and the New SOLO Mode Experience?

ByteDance's AI programming assistant TRAE has officially released version 2.0 and is gradually opening access. This version introduces SOLO mode - a system with context engineering capabilities that can autonomously understand requirements, decompose tasks, call tools, and complete development tasks from requirement analysis to final deployment.

From TRAE Core Developer: The Journey from MarsCode to TRAE

Tianzhu (Core TRAE Developer):

Thanks for the invitation. Interest disclosure: I'm a core TRAE developer. Without realizing it, I've been working in this field for half a year. Today I want to share my thoughts on "Rebirth: Working on AI Coding at a Big Tech Company."

Timeline Overview

The timeline is quite simple:

• November 15, 2022: An unexpected severance package made me rethink my future plans. Fortunately, I soon joined ByteDance.
• November 15, 2023: I was called to a retreat in Hangzhou to lead the MarsCode Cloud IDE team and related work. It was a year of crazy running.
• November 15, 2024: Entered another retreat for six intense months, resulting in the TRAE 1.0 and 2.0 evolution everyone sees today.

Half a year has passed since then. During the Force conference last month, Dingkun shared in his keynote "Talking about AI Coding" that TRAE's monthly active users have exceeded one million. We've delivered a milestone achievement.

From MarsCode to TRAE

Those familiar with us know that throughout 2024, we released many Cloud IDE related products, including MarsCode Cloud IDE, Coze IDE, Juejin AI coding practice, and our internal ByteDance Cloud IDE, which has quite high user coverage.

First, let me clarify a concept: Cloud IDE ≠ Web IDE. From my perspective, IDE consists of frontend interaction layer and business logic layer:

• Frontend interaction layer can run in browsers or local Electron, can be full-featured or lightweight
• Business logic layer can run locally, in remote K8S containers, or in browser WebContainer

Cloud means "Code Anywhere" - the interaction and logic layers can be freely combined, whether on the same machine, separated through SSH Remote connection, or even used through iPad. This is my understanding of Cloud IDE.

Performance Achievements

In 2024, we invested tremendous effort in optimization, not only extensively Rustifying VSCode but also deeply customizing at the K8S level. We achieved excellent results in performance, cost, and stability. Our end-to-end startup performance P90 reached 5 seconds, while:

• GitHub Codespace needs 30 seconds
• Google IDX needs 1 minute

We've achieved world-class performance on this metric.

My Understanding of AI Coding

LLM models essentially predict the next character. Compared to complex natural language, programming languages are more concise, rigorous, and predictable. Therefore, AI Coding has become the first PMF product in this wave.

From my perspective, AI Coding is similar to autonomous driving, with several stages:

AI-Assisted Programming → AI Pair Programming → AI-Driven Programming

Currently, TRAE Builder, Cursor Composer, and Windsurf Cascade are all targeting the AI Pair Programming stage.

AI-Assisted Programming

Code Completion Evolution:

• Early stage: Code suggestions through dropdown lists to select object methods and properties
• ChatGPT era: Copilot's breakthrough with Ghost Text interaction - just one Tab to accept suggestions
• Advanced stage: Super Completion (Tab Tab) evolution from "predicting next character → predicting next edit position" and "new code → modifying existing code"

Code Generation Development:

• Initial shock: Conversational generation of algorithms or pages in ChatGPT
• IDE integration: SideChat plugins like Copilot Chat, Codeium with direct IDE context access
• Technical challenges: Project understanding, context clipping, model capability, and engineering PE capability
• Solution evolution: Fast Apply code merging - automatically generating and merging code to appropriate locations

AI Pair Programming

What is AI Agent?

As AI capabilities continue growing, user demands become more complex. We expect AI to have more autonomy, like a high-potential intern pair programming with us - the community calls this AI Agent.

Agent's core capabilities:

• Thinking and scheduling abilities
• Context environment awareness
• Tool invocation capabilities

Trae Agent Evolution:

Agent 1.0 (Initiated December 17, 2024):

• 20 days of intense development
• Workflow-dependent: Think → Plan → Execute → Observe cycles
• Limited by LLM capabilities at the time

Agent 2.0 (Released April 8, 2025):

• 21 days of redesign and development
• Greater LLM autonomy for active requirement understanding
• Enhanced environment perception and tool-driven execution
• Significantly improved LLM capability utilization

Trae SOLO Mode

As TRAE continued evolving, we gradually realized AI and IDE's dominant relationship is reversing:

• Previously: Programmers code in IDE, AI assists
• Now: AI dominates coding, IDE/Browser/Terminal are just its tools

Therefore, we explored SOLO mode, upgrading from "AI assists IDE" to "AI dominates IDE" collaboration mode.

SOLO Mode Features

This 2.0 release mainly provides:

• SOLO: AI-centric interaction mode
• Built-in end-to-end delivery Agent for Web APP scenarios
• Future expansion: More specialized Agents planned

Community User Experiences

Real-World Testing by Orange90 (Juyige)

"I got an invitation code from the TRAE team and tried SOLO mode. My evaluation: TRAE SOLO currently outperforms v0, nocode, and similar products."

Testing Process:

1. Simple Interface: No project setup needed, just start with requirements
2. Requirement Input: "Help me generate a fully functional application that renders Mermaid code into images, with handwritten fonts (supporting both Chinese and English), and deploy it on Vercel."
3. Automated Workflow:
   • Project folder creation prompts
   • Automatic PRD (Product Requirement Document) generation
   • Dependency installation and environment setup
   • Problem-solving and project completion
   • Live preview and Vercel deployment authorization

Comprehensive Testing by Xu Fei

Tested three complex scenarios with TRAE 2.0 SOLO mode:

• Task board management system ✅ Successful deployment
• Simple procurement contract management system ✅ Functional implementation
• Workflow editor ⚠️ Partial success with some limitations

Results: Only 3-4 requirement inputs plus debugging communication (total ~10 interactions) achieved professional-quality results that would take significant human development time.

Developer Perspective from Death Moon Scarlet (TRAE Team)

"Interest disclosure: I'm a TRAE development team programmer."

Challenge: Often have implementation ideas but struggle with execution due to skill gaps, detail control issues, and incomplete project completion.

Success Story: Used TRAE SOLO to recreate MD5 Battle - a popular Flash game from 2008-2009.

Simple Requirement Given:

Architecture Insight: TRAE SOLO implements Multi-Agent architecture with specialized requirement Agent, development Agent, etc.

How to Properly Collaborate with AI Interns?

I've always defined AI as: High-potential interns that everyone can have.

Many people find AI programming either amazing or useless - this often comes from not finding the right way to work with interns at different stages, not managing expectations properly.

My approach: Treat current AI interns like real human interns:

• Assign appropriate tasks
• Provide suitable guidance and information input
• Be ready to provide backup and take over when needed
• Manage expectations properly

Common Mistakes:

• Hands-off approach: Not providing sufficient context, expecting AI to complete formal projects with one-sentence requirements
• Beyond capability boundaries: Expecting code-writing AI to generate inappropriate content
• Low hiring standards: Wanting Claude 4 capabilities while only willing to pay for OpenAI 3.5 free
• Unrealistic ROI expectations: Expecting $2000 token usage output from $20 monthly fees

Conclusion and Future Outlook

When we first started the project, our TL told us: "We have willingness, capability, opportunity, ideas, and support. I firmly believe we can create a world-class product." I initially took it as "vision talk."

But when doing year-end review in late 2024, looking at MarsCode's technical metrics, and in mid-2025 seeing TRAE's million monthly active users and our team's accumulated insights, I suddenly realized we're actually steadily running on this path. I truly believe now.

Following the philosophy of "research one generation, deploy one generation, share one generation," we're implementing Trae Agent 3.0 architecture to support Multi Agent and Remote Agent explorations.

The road is long and winding, but we'll keep exploring. Professional productivity tool disruption will inevitably reshape developers' cognition and development methods. Future IDEs may no longer be "code-centric" as current ones - this transformation might happen within 3 years.

---

Article Source Information

• Original Title: 如何评价 Trae 发布 2.0，以及新推出的 SOLO 模式体验如何？
• Source Platform: Zhihu Q&A Community
• Publication Date: July 22, 2024
• Contributors: Core TRAE developers and community beta testers
• Reprint Note: This article compiles authentic developer community discussions and experiences from Zhihu Q&A platform, translated for international readers