Supported agents

Supported agents¶

Every agent in the catalogue, the ports each version has proven, and the benchmarks it fits.

Agent	Versions (bare key: the default)	What it is	Driven over	From	Licence
Codex	`codex@2.2.2`, `codex@0.16.0`	OpenAI's coding agent.	ACP via `codex-acp`	npm	Apache-2.0
Goose	`goose`	The AAIF's open-source agent, formerly Block's.	ACP	npm	Apache-2.0
JiuwenSwarm	`jiuwenswarm`	openJiuwen's agent.	ACP	PyPI	Apache-2.0
mini-swe-agent	`mini-swe-agent`	Princeton's 100-line agent, bash only.	a command	PyPI	MIT
OpenCode	`opencode`	The open-source terminal coding agent.	ACP	npm	MIT
Terminus 2	`terminus-2`	Harbor's terminal agent, used by Terminal-Bench.	a command	PyPI	Apache-2.0

ACP (start.acp) can prove task, workspace, tools and conversation; a command (start.command with {prompt}) task and workspace only.
Examples, never offered for a run: reference (Example agent: Nullius's example of the protocol, to read or copy.)

Port	Label	What it proves
`agent.task/v1`	One task	One prompt in, one turn; the answer is the agent's final message.
`agent.workspace/v1`	Workspace files	The session's working directory is the unit's workspace; the agent reads and writes files there.
`agent.tools/v1`	MCP tools	The agent calls the tools of the MCP servers given to its session.
`agent.conversation/v1`	Conversation	Several prompts in one session, each answered in its own turn, the earlier ones remembered.
`agent.browser/v1`	Browser	A browser reached through an MCP server Nullius starts.

Agent	Ports				Tools, the skills of `agent.tools/v1`
Agent	One task	Workspace files	MCP tools	Conversation	decoys	chain	retry
`codex@2.2.2`	✓	✓	✓	✓	✓	✓	✓
`codex@0.16.0`	✓	✓	✓	✓	✓	✓	✓
`goose`	✓	✓	✓	✓	✓	✓	✓
`jiuwenswarm`	✓	✓	✓	✓	✓	✓	✓
`mini-swe-agent`	–	✓	–	–	–	–	–
`opencode`	✓	✓	✓	✓	✓	✓	✓
`reference`	✓	✓	✓	✓	✓	✓	✓
`terminus-2`	–	✓	–	–	–	–	–

✓ proven by cli conform, – not proven. An agent that misses a skill still has the port. Browser has no kit yet, so no agent is tested on it; it comes with the first benchmark that needs it.

	GSM8K	HumanEval	SWE-bench Verified	τ³-bench	Terminal-Bench 2.0
	One task	Workspace files	Workspace files	MCP tools, Conversation	Workspace files
`codex@2.2.2`	✓	✓	✓	✓	✓
`codex@0.16.0`	✓	✓	✓	✓	✓
`goose`	✓	✓	✓	✓	✓
`jiuwenswarm`	✓	✓	✓	✓	✓
`mini-swe-agent`	✗	✗	✓	✗	✗
`opencode`	✓	✓	✓	✓	✓
`reference`	✓	✓	✓	✓	✓
`terminus-2`	✗	✗	✗	✗	✓

Every agent harbor agents list finds in the pinned Harbor, as Harbor built-in, and the builder's own:

Agent	Kind	What it does
`tau3-llm-agent`	reference	τ³'s own reference agent, from Harbor's source tree.
`oracle`	control	Applies the task's reference solution; never calls a model. A ceiling, never ranked.
`nop`	control	Does nothing. The floor a verifier must reject.