Agent Harness
Layers
Every agent harness implements the same five extension layers — only the names differ, and occasionally conflict. Instructions persist, skills teach, MCP connects, bundles ship, subagents delegate. Tap a layer heading for detail.
Where to start
Start from what you're trying to change, not from the feature name.
The five layers, explained
Same architecture everywhere, and the order is the stack: rules → know-how → access → packaging → parallelism. Each layer only earns its place when the one before it isn't enough.
Terminology traps
- "Plugin" has several meanings across these tools. Claude, Codex, Goose, Grok Build, Cursor, and Copilot use it for installable bundles. Hermes and OpenCode use it for executable code extensions. OpenClaw supports both native code plugins and imported content bundles, while Osaurus imports Claude plugins into its own feature set. Check the manifest and payload, not just the name.
- "Extension" flips meaning too. In Goose, an extension is an MCP server (tool access). In Gemini CLI, an extension is a full agent bundle. In Zed it is an editor bundle that can carry an MCP server but has no documented skill or instruction payload. Do not confuse current Copilot plugins with the retired Copilot Extensions GitHub Apps.
- "Skills" share a core format. Every harness here uses the same Agent Skills / SKILL.md core format. The wrinkles are discovery paths, precedence, trust, and distribution: several tools read Claude-compatible and vendor-neutral
.agents/skillsdirectories alongside their own. - "Connector" is mostly Claude-speak for a packaged remote MCP server — though the word shows up generically elsewhere to mean any MCP integration.
- "Bot" is not a subagent. A bot is a named teammate that stays. Hermes Bot Mode is a roster over profiles; Osaurus Agents are saved configs; Grok Bot is a separate always-on product from Grok Build. This column is a throwaway worker: spawn, return a summary, done.
The layer underneath: MCP
The Model Context Protocol (open standard, released by Anthropic in Nov 2024 and now governed by the Linux Foundation's Agentic AI Foundation) is the shared plumbing for the "tool access" row — a JSON-RPC protocol where a server publishes tools any compatible agent can discover and call. This is why the ecosystem interoperates: a Goose extension, a Claude connector, a Hermes MCP entry, and an MCP server inside a Codex plugin can be the exact same server. Osaurus demonstrates the two-way nature of the standard: it consumes remote MCP servers and exposes its own local tools as an MCP server that other harnesses can connect to — so your Osaurus toolbox can serve Claude Desktop. Skills (SKILL.md / agentskills.io) and instruction files (AGENTS.md) are similarly portable open standards, and since October 2026 all three sit under that same vendor-neutral foundation, alongside goose. The names are vendor branding; the files and protocols underneath increasingly are not.