LAGO reuses specialist agents to solve coding tasks
In plain words
When one computer agent handles a big coding task, it has to plan, search files, make changes, test them, and recover from mistakes. LAGO is a system that lets a coordinating agent hand parts of a task to specialists and reuse successful specialists later.
On the same 150 coding issues, LAGO solved more than two other systems rerun with the same model. The comparison did not match their context windows or computing budgets, so it cannot show which part of LAGO made the difference.
The research
When a coding task takes many steps, one agent must plan, find relevant context, edit files, test its changes, and recover from errors. Could a system reuse specialists for recurring subtasks while still checking their work? That is the question LAGO addresses.
LAGO assigns subtasks to reusable or newly created specialists, then checks their work and keeps successful additions in a library.
The paper asks whether agents can handle complex tasks without repeating familiar subtasks or relying on one long tool-use chain. LAGO lets a coordinating agent retrieve or create specialist agents, check their work, and reuse successful specialists across tasks.
On 150 shared code-repair issues, LAGO resolved more issues than two other systems rerun with the same model and evaluator. This suggests the full LAGO scaffold helped in that comparison, though the test cannot identify which component was responsible.
The shared comparison covers 150 tasks, and the paper estimates about ±8 percentage points of binomial uncertainty at that size. Its smaller component tests are noisy and do not fully separate the value of reusable skills from the effect of the prompts that come with them.
Analysis
Deep
- Original paper authors
- Yuran Wu, Mengkang Hu, Qionglin Qiu, Fanqing Meng, Qiguang Chen
- Submitted to arXiv
- License
- CC BY 4.0