Agent Toolbench: better hands for coding agents on Windows

[puntastic]

Better hands for coding agents. Small, runnable experiments on the boundary between an agent's decision and the tools that carry it out.

We started with “would Bash be easier for a coding agent than PowerShell?” and found a more useful seam: sometimes the requested Bash wasn't being launched at all. Codex could silently fall back to CMD.

That led to a small Windows execution lab:

a candidate fix for shell discovery and explicit failure reporting;
an additive argv route for ordinary executable/argument calls;
a runnable comparison of direct execution, PowerShell, Git Bash, and Nushell;
results, negative controls, regression tests, and the limits of what they show.
The fixed transport cases passed with direct arguments, Codex-style UTF-8 PowerShell, scoped-literal Bash, and Nushell. Plain PowerShell and default Git Bash each exposed a different edge on our machine. This is not a claim that one shell is universally best, or that we have measured better model reasoning.

The interesting direction is making the interface do more of the mechanical work while retaining capable shells and the existing execution machinery.

Agent Toolbench

You can read the findings without installing anything, reproduce the comparison, or inspect the patches. Reproductions and counterexamples are welcome. The work is AI-assisted, independently maintained, and currently Windows-specific.

[bomtobe-hgkim]

A useful negative control might be a case where shell discovery succeeds but PATH, working directory, or UTF-8 handling changes the executable that actually runs. Reporting the requested shell, resolved executable, cwd, argv/encoding, exit code, and fallback reason together would make a silent CMD fallback distinguishable from success in the requested shell. I work with BOMTOBE on Orke, a Claude Code/Codex desktop workspace; I’m on Linux and have not reproduced this Windows-specific setup, so this is a test suggestion rather than a result.