agent-tool
The Problem
Claude Code turns are expensive to spend on writing code. I wanted a way to keep Claude focused on planning and review, while the bulk of the actual coding happened somewhere cheaper and ran in parallel instead of one turn at a time.
What I Built
A small CLI orchestrator. Claude Code decomposes a task into an agent roster and writes a plan file; agent-tool then launches Antigravity CLI (agy) and GitHub Copilot CLI (copilot) as subprocesses, each in its own isolated git worktree, so they can work at the same time without stepping on each other. A separate Textual TUI tails the whole run live in another terminal, without touching Claude at all.
It also supports tiered, dependency-ordered plans—architects who set constraints, implementers who write the code, verification roles that audit another agent’s actual work against a real verify_command. Everything after the plan file is written is decoupled through plain files under .agent-tool/runs/<run_id>/—no daemon, no extra API key beyond what the CLIs already cost.
Why It Matters
Coding agents lying about their own success is a real, recurring failure mode—I’ve seen an LLM “verify” Kotlin that didn’t compile by reading the diff instead of building it. Prompting harder doesn’t fix that; a real subprocess with a real exit code does. That’s the idea agent-tool is built around: let the model plan and review, but make anything claiming success actually prove it.