Mandip Adhikari
← Index
Project2025

Kavi

A CLI tool that runs multiple AI coding agents (Codex + Claude Code) in parallel and solves the coordination problem that emerges the moment more than one agent can touch a codebase.

TypeScript · Node.js · CLI · Git worktrees · Multi-agent · Process management · JSON persistence

2+
agents orchestrated in parallel
per-task
git-worktree branch isolation
every action
approval-gated

The problem

One coding agent is a tool. Two coding agents touching the same repository is a distributed-systems problem — races on the working tree, conflicting edits, and no shared memory of what was already tried.

Kavi treats agent orchestration as exactly that: isolated workspaces, an approval boundary on every mutation, and a persistent knowledge base so context survives across sessions.

Architecture

A central orchestrator dispatches tasks to agents, each running in its own git worktree so their changes never collide. A persistent Brain carries knowledge across runs; shadow missions let strategies compete before anything lands.

            ┌──────────────────────────────┐
   task ───▶│         orchestrator         │
            └──────┬───────────────┬───────┘
                   │               │
          ┌────────▼──────┐ ┌──────▼────────┐
          │  Codex agent  │ │  Claude Code  │  ◀─ approval gate
          └────────┬──────┘ └──────┬────────┘     on every action
                   │               │
          worktree/task-a   worktree/task-b
                   │               │
            ┌──────▼───────────────▼──────┐
            │   Brain — persistent KB      │  carried across sessions
            └──────────────────────────────┘

What it does

Maintains a persistent Brain knowledge base across sessions so agents don't relitigate solved problems.

Enforces approval gates on every agentic action — nothing mutates state without a human checkpoint.

Runs shadow missions to compare strategies head-to-head, and isolates each task on its own worktree-based branch.