MarsCode Agent is a framework from ByteDance that uses LLMs to automatically identify and repair bugs in software code. It is introduced in MarsCode Agent: AI-native Automated Bug Fixing, which describes it as building an agent framework and providing agents with interactive interfaces and tools for code retrieval, debugging and editing, so that agents can take over some software engineering tasks.
Capabilities
Work is divided among six roles. The Searcher collects code snippets related to the issue using the code knowledge graph and the language server protocol; the Planner analyzes them and classifies the issue into a dynamic debugging or a static repair workflow; the Reproducer writes a reproduction script and confirms the issue reproduces in a sandbox; the Programmer edits the code; the Tester checks each version against the reproduction script; and the Editor proposes multiple repair solutions for static repair. Only the Programmer and the Editor can edit code, only the Programmer can reset the repository, and only the Reproducer and the Tester can run reproduction scripts. In dynamic debugging the Programmer and the Tester iterate until the issue is resolved, in a runtime sandbox set up in a Docker container. In static repair, drawing on an approach similar to Agentless, several candidate fixes are generated in a single LLM request, normalized with an AST, merged and voted on.
For retrieval, MarsCode Agent builds a code knowledge graph whose nodes are code entities such as variables, functions, classes and files, and whose edges record file structure, function calls and symbol references; an agent’s query goes through entity recognition, embedding similarity and keyword search against the graph, and the merged candidates are re-ranked. The language server protocol complements the graph for definitions and references outside the target project, with fuzzy positioning that works out an LSP request from a file name, line number or identifier the agent supplies. General file search and grep are also available.
Code edits are written in AutoDiff, a format inspired by Aider that resembles git conflict markers: the agent gives the file path, the original code and the replacement code, and the tool matches the original snippet to the most similar segment of the file, replaces it, adjusts indentation and produces a unified diff. Each edit is then checked with LSP static diagnostics before and after the change, and if new errors at the Fatal or Error level appear, the diagnostics are returned to the agent for further changes.
Adoption & Ecosystem
The report evaluates MarsCode Agent on SWE-bench Lite, part of SWE-Bench, and compares its code retrieval with other publicly traceable solutions, including CodeR, Moatless and Agentless. Its authors list reducing LLM call costs, improving user-agent collaboration and supporting dynamic debugging within real user workspaces among their plans for the framework.