I don’t believe agents should “think” without being allowed to check their work. So I built a CLI agent that can actually run code while it reasons.

Subconscious handles the reasoning and tool-calling while actively pruning context as it goes, so the agent runs faster, costs less, and has more tokens available for complex tasks.

Code runs in one of the fastest sandboxed environments available using E2B, keeping execution quick while your local machine stays clean, organized, and protected from prompt injection.

The result feels less like a language model and more like working with a tool that can verify its own work, stay disciplined, and generate real reports and charts.

Run it with:

npx @subcon/e2b-cli

If you care about reliable agents, CLIs, or running code safely without touching your local machine, this is worth a look. Especially if you are tired of deleting random .md files.