Multi-Tool AI Coordination
Multi-agent isn't the problem. Multi-tool is.
Inside a single environment like Claude Code, agents coordinate fine. They share orchestration and context. Run 20 subagents and they stay aligned because the tool manages that coordination for you.
The trouble starts when you run multiple AI tools on the same repo. Claude Code for refactors. Codex CLI for tests. Gemini CLI for research. Each tool operating with its own memory, its own assumptions, and no shared state between them.
I ran Claude Code and Codex CLI on the same codebase. Claude refactored a module. Codex updated tests for that same module against the pre-refactor interface. Both saved their changes. The tests failed. Neither tool knew the other existed.
The Program Management Insight
Multi-tool AI coordination maps directly to enterprise program management. I know because I spent 23 years solving the human version.
| Human Team Failure | Multi-Tool AI Failure |
|---|---|
| Developers overwrite each other's work | Tools edit the same file simultaneously |
| Knowledge gets lost in handoffs | Decisions disappear between tools |
| No visibility into cross-team progress | No dashboard across tools |
| Governance enforced at audit time | Quality checked at PR time |
| Context stays siloed | Each tool starts from scratch |
How Forge Coordinates Multiple AI Tools
Single state directory, not a central server. The .forge/ directory in your project root contains task state, file locks, knowledge entries, and governance configuration. All three AI tools read from and write to this directory. No daemon. No database. No network calls. The filesystem is the coordination layer.
Three native adapters. Claude Code communicates via MCP stdio. Codex CLI and Gemini CLI use filesystem conventions. No tool needs to know about the others. They all see the same state through their native interface.
File locking that actually works. When Claude Code claims src/auth.ts, the lock is recorded in .forge/locks/. Codex CLI checks before editing and queues if the file is locked. The lock manager is built into the Rust core. 292 tests cover concurrent access patterns, timeout behavior, and deadlock prevention.
Knowledge that compounds across tools. Decisions made during a Claude Code session are captured in .forge/knowledge/. When Codex CLI starts the next day, those decisions are available. The knowledge flywheel works across tools, not just within a single tool's session.
Task board with dependency awareness. Forge tracks which tool is working on what, which tasks depend on which, and which are blocked. Tools claim available tasks, report progress, and check for blockers. All through the shared state directory.
Architecture
The Rust orchestrator binary (4MB, zero runtime dependencies) is the coordination core. Policy enforced here. Nowhere else. The plugin and UI are adapter surfaces. Communication uses .forge/ filesystem and MCP stdio. Any tool that can read files and make MCP calls can participate in Forge's coordination. The architecture is extensible by design.