Agent Harness Bundle
AI coding infrastructure
Long coding sessions lose context, and parallel agents can overwrite each other's work. I built a shared toolkit that gives coding agents common rules, tasks, and file access.
The problem
A single coding agent can lose the important rules as the session grows. Multiple agents create a second problem: two panes can edit the same file, repeat work, or accept a change that fails the project checks.
What I built
The harness keeps one shared rules directory, procedural memory, dependency-aware bead tracking, Agent Mail reservations, and command guards. Claude Code, Grok Build, Antigravity, and Kimi can read the same operating context through thin adapters.
Engineering decisions
The system separates durable rules from model-specific adapters. File reservations are advisory until a commit guard checks them, task graphs make dependencies explicit, and destructive commands require a deliberate safety path.
Why it matters
The harness records the task, reservation, and commit checks needed to review a change. Reservations stay advisory until the pre-commit guard checks active leases.