{"id":16,"date":"2025-01-02T05:48:00","date_gmt":"2025-01-02T05:48:00","guid":{"rendered":"https:\/\/codegenblog.kinsta.cloud\/?p=16"},"modified":"2025-08-27T12:10:18","modified_gmt":"2025-08-27T12:10:18","slug":"act-via-code","status":"publish","type":"post","link":"https:\/\/codegen.com\/blog\/act-via-code\/","title":{"rendered":"Act via Code"},"content":{"rendered":"\n

Two and a half years since the launch of the GPT-3 API, code assistants have emerged as potentially the premier use case of LLMs. The rapid adoption of AI-powered IDEs and prototype builders isn\u2019t surprising\u2014code is structured, deterministic, and rich with patterns, making it an ideal domain for machine learning. Developers actively working with tools like Cursor (myself included) have an exhilarating yet uncertain sense that the field of software engineering is approaching an inflection point.<\/p>\n\n\n\n

Yet there\u2019s a striking gap between understanding and action for today\u2019s code assistants. When provided proper context, frontier LLMs can analyze massive enterprise codebases and propose practical paths towards sophisticated, large-scale improvements. But implementing changes that impact more than a small set of files with modern AI assistants is fundamentally infeasible. The good news is that for focused, file-level changes, we\u2019ve found real success: AI-powered IDEs (Windsurf<\/a>, Cursor<\/a>) are transforming how developers write and review code, while chat-based assistants are revolutionizing how we bootstrap and prototype new applications (via tools like v0<\/a>, lovable.dev<\/a>, and bolt.new<\/a>).<\/p>\n\n\n\n

However, there\u2019s a whole class of critical engineering tasks that remain out of reach\u2014tasks that are fundamentally programmatic and deal with codebase structure at scale. Much of modern engineering effort is directed towards eliminating tech debt, managing migrations, analyzing dependency graphs, enforcing type coverage, and other global concerns. Today\u2019s AI assistants can propose solutions but lack the mechanisms to execute them. The intelligence is there, but it\u2019s trapped in your IDE\u2019s text completion window.<\/p>\n\n\n\n

The bottleneck isn\u2019t intelligence\u2014it\u2019s tooling. The solution is giving AI systems the ability to programmatically interact with codebases through code execution environments. These environments are the most expressive tools we can offer agents, enabling composition, abstraction, and systematic manipulation of complex systems. By combining code execution with custom APIs for large-scale operations, we unlock new high-value use cases.<\/p>\n\n\n\n

Beating Minecraft with Code Execution<\/strong><\/h2>\n\n\n\n

In mid-2023, a research project called Voyager<\/a> solved Minecraft, performing several multiples better than prior SOTA. This success wasn\u2019t about raw intelligence\u2014it was about providing a more expressive action space: code.<\/p>\n\n\n\n

GPT-4, when allowed to write and execute JavaScript programs through a clean API, could craft high-level behaviors and reuse learned \u201caction programs\u201d across tasks. This enabled skill accumulation, experience recall, and systematic reuse.<\/p>\n\n\n\n

\n

\u201cWe opt to use code as the action space instead of low-level motor commands because programs can naturally represent temporally extended and compositional actions\u2026\u201d<\/p>\n<\/blockquote>\n\n\n\n

Code is an Ideal Action Space<\/strong><\/h2>\n\n\n\n

Letting AI act through code rather than atomic commands yields a step change in capability. In software engineering, this means expressing assistant behavior through code that manipulates codebases.<\/p>\n\n\n\n

# Implement `grep` via for loops and if statements\nfor function in codebase.functions:\n    if 'Page' in function.name:\n        function.move_to_file('\/pages\/' + function.name + '.tsx')<\/code><\/pre>\n\n\n\n
This paradigm brings multiple advantages:<\/p>\n\n\n\n