An agent that generates and improves software from natural language instructions: it lets you specify software in natural language, watches as an AI writes and executes the code, and can implement improvements on request.
Generating new code from natural-language prompts and improving existing code based on instructions. It also installs a 'bench' binary for benchmarking custom agent implementations against public datasets (APPS, MBPP), supports vision-capable model image inputs, and supports OpenAI, Azure OpenAI, Anthropic, and open source/local models.
The README explicitly describes an autonomous agentic workflow: "Sit back and watch as an AI writes and executes the code." The agent generates code from a natural-language prompt and then runs it, with code-surface evidence confirming shell command execution and filesystem writes. There is no evidence of a code-enforced per-step confirmation gate before generating, writing, and executing code — actions proceed as a multi-step loop. This places it at Tier 4 (acts autonomously). It is not Tier 5, as there is no evidence of self-modification of its own code/prompts or spawning of new agents.