Why AI Agent Orchestration Needs Its Own Language

Something has been bugging me about the AI agent space: there's no standard way to define a team of agents.

The Problem

Every platform has its own config format. Claude Code wants AGENT.md files with YAML frontmatter. Cursor wants .mdc rules. OpenClaw wants soul.md plus skill files plus channel configs plus gateway setup. Codex wants TOML. And so on.

So if you build a 5-agent pipeline — say a code review system with an analyzer, security scanner, reviewer, reporter, and an approval gate — you end up with 20+ config files scattered across directories. No overview, no validation, no portability. Pure vibe coding.

Then someone says "we need this in Cursor too" — and you rewrite the whole thing from scratch. Different format, different structure, same agents.

This felt like the exact same problem infrastructure had before Terraform. Every cloud provider wanted different configs. No portability. No single source of truth.

What I Built

I built AgenTopology — a declarative language specifically for multi-agent systems. You define agents, flows, quality gates, hooks, group chats, and MCP servers in one readable file. Then one command scaffolds it to whichever platform you use.

Currently supports 7 targets: Claude Code, OpenClaw, Cursor, Codex, Gemini CLI, Copilot, and Kiro.

The Part That Makes It Actually Usable

It ships with a Claude Code skill. You type /agentopology, describe what you want — "give me a code review team with an analyzer, scanner, and reviewer" — and it builds the whole thing. No syntax to learn. No docs to read.

Then you can visualize the entire topology as an interactive graph. Click any node and see all its tools, connections, MCP servers, hooks, and config. This is what a code review pipeline looks like in the visualizer:

Every agent, every connection, every gate — visible in one view.

How It Works

1. Describe your team (via skill or write an .at file)
2. Validate — 29 built-in rules catch errors before you deploy
3. Visualize — interactive graph of the full architecture
4. Scaffold — one command generates all platform configs

Want to switch from Claude Code to OpenClaw? Change --target claude-code to --target openclaw. Same team, zero rewrite.

Want to add a hook? Tell the skill "add a hook that formats code after every write." Want to add a gate? "Add a quality gate between builder and reviewer." The skill handles the syntax.

Try It

npm install -g agentopology

Then in Claude Code:

ln -s $(npm root -g)/agentopology/skill .claude/skills/agentopology

Type /agentopology and describe your team.

Open source, Apache 2.0: github.com/agentopology/agentopology

Website: agentopology.com

---

I'm curious — is a declarative approach the right direction for multi-agent orchestration? Or is it too early to standardize? Would love to hear what problems you've run into building agent teams.