Self-healing code from Docs

The idea is basic: generate python code based on library documentation, on which the LLM hasn’t been trained. But with the current state of LLMs, this turns out to be overly optimistic. So we relocate the effort of the task to the environment engineering. How can the environment, with which the agent interacts, maximally contribute to the agent’s evolution and success? Fortunately, the task of code generation offers a rich feedback loop, from static code analysis tools, to run time execution logs, to code base knowledge graphs. In this talk I will make a demo of the coding assistant and provide a sneak peak into the ongoing process of enriching the environment/agent interaction for optimal agent performance.