TRAE Merges Code and Work Into Unified Platform With Agent and IDE Modes

In the old setup, a user modifying a bug would open TraeCode, while producing a technical plan required switching to TraeWork. The new TRAE places Agent mode as a global workbench where a project can be managed in one window and different jobs can be assigned to multiple agents for coding, documentation and delivery preparation. IDE mode retains SOLO and IDE options, with a switch in the upper-right corner.

QuantumBit described a trial in which three agents worked in the same project folder on a 3D pelican riding a bicycle. One agent wrote the code for the pelican, bicycle and scene; a second acted as tester, opening the page, taking screenshots, checking results and listing problems; a third fixed the problems on that list. The agents worked in relay, and test reports, repair reports and delivery notes were saved automatically under My Artifacts and could be shared directly.

The report also tested whether a non-programmer could turn an idea into a game. In a project for a 3D dungeon map editor, two tasks ran at the same time: one agent wrote a game design document, producing a Word plan and a Markdown version, while another produced a high-fidelity interface design covering edit mode and first-person play mode in a low-poly dungeon style. A third task then read both outputs and wrote code.

Because the planning and design agents worked in parallel and could not see each other's output, they produced conflicts. The grid was 20 by 20 in the design but 32 by 32 in the plan; the eraser shortcut was E in one and X in the other; monster spawn points were set at one in the plan and three in the design. The development agent read both files, chose a simple playable version as its principle, resolved the conflicts and recorded the reasons in the delivery note. After coding, the development agent opened a browser to test the result and fixed six issues, including a paused window that appeared on entering play mode and a play screen too dark to see corridors. The final editor allowed users to place floors, walls, doors, chests, levers and monster spawn points, and included undo, redo, autosave and JSON import and export.

The report noted that Notch, the creator of Minecraft, has been working on a new game called Levers and Chests, and the editor's chests and levers were described as a small tribute. In IDE mode, QuantumBit said it asked the system to adjust first-person walking speed to 70 percent. The speed dropped from 3.4 tiles per second to 2.4 tiles per second, and running speed was reduced in step. The report described Agent mode as useful for letting AI handle complex development while IDE mode is for reviewing and adjusting details line by line.

For professional developers, QuantumBit tested a simulated project called TeamDesk, a small task management web page with about 20 files and a fictional data set plus an issue list. Three agents were assigned at once: one read the code and produced an architecture explanation for newcomers; one fixed bugs without touching test files; one added regression tests without touching business code. The architecture agent delivered a web report explaining the project's layers, how page actions changed data and where data was stored in the browser. The bug-fixing agent reproduced the issue from the list, found the cause and changed one line to exclude archived tasks from statistics. The test agent added regression coverage. The report said the business code had only one line changed and the test folder gained one new file. Regression tests could also be scheduled to run every morning, with alerts when problems appear.

In another scenario, QuantumBit built a feedback handling console called FeedbackDesk for an internal review. The work required a product requirements document, an operable web demo and a review presentation. Three agents again worked at once: one wrote the PRD, one built the web demo and one prepared the presentation. They produced a four-page Word requirements document, a browser-based demo and a six-page presentation with a script. QuantumBit said the same work previously required TraeWork for documents and presentations and TraeCode for the web page, while the new TRAE advanced them in one window. The platform also offers a template library for document work. QuantumBit added that after binding Feishu or WeChat, users can mention TRAE to handle cross-platform work even outside TRAE, and the mobile client stays synchronized.

Asked why the two products were merged, the report said coding is only one part of development. A project moves from an idea to delivery through requirement breakdown, proposals, design, testing, bug fixing and reporting. Those steps were split between two tools, breaking context and scattering artifacts. New TRAE connects the chain into a full-pipeline development platform. Vertically, documents, design, code and tests can live in one project, repetitive processes can be automated, and desktop, web and mobile remain synchronized. Horizontally, one workbench manages all projects, multiple agents can run in parallel, the left sidebar shows where tasks are stuck, and users can switch back to IDE for deep code changes.

The report noted that TRAE is not alone in this direction. In April, Cursor released Cursor 3, adding an Agents Window for managing agents while keeping the classic IDE interface, and users can switch between the two. Claude Code also added Agent View, which can start, manage and monitor multiple background agents in one panel. According to QuantumBit, leading overseas AI coding tools are putting agent management and code editing into the same tool.

The tests also showed that more agents do not automatically improve results. In the pelican example, the development agent said it had tested twice in the browser and everything was normal, but the testing agent found four problems, including reversed arrow keys: pressing right sent the pelican left. The dungeon editor's planning and design agents contradicted each other, while FeedbackDesk produced three outputs with matching details. QuantumBit concluded that the human role shifts toward breaking work down clearly, preparing a shared basis for agents, and deciding who does the work and who checks it.

The report said TRAE, described as an established player in AI coding, aimed to use one workbench to cover each stage of development and serve more developers who want to build. It framed the update as moving AI coding from having one agent write code to directing a group of agents from idea to delivery, a phase it called Agent development.

Editor's Summary

TRAE has merged its Code and Work products into a single platform with Agent and IDE modes, allowing users to manage multiple agents, share artifacts and move from requirements to delivery in one place. QuantumBit's tests showed benefits in multi-agent coding, documentation, testing and regression work, while also revealing coordination problems when agents run in parallel. The update reflects a broader shift among AI coding tools toward combining agent management with code editing.