Reference

Agents

hbench's built-in agents, the eight the site issues runs of, every option you can set, the options a board holds, models and keys, and prompt templates.

An agent is the harness that turns the model's replies into commands and shows the model the results. Any of hbench's built-in agents can go on a board, with any options, and so can an ACP agent of your own. The list at the end of this page has every built-in agent and its options.

Common agents#

The site issues runs of these 8 agents. Upload a job to put any other built-in agent on a board.

Set options#

In Run a benchmark, an agent's options are the fields under More options, in Agent settings; leave one blank for the agent's default. With hbench, pass each as --ak NAME=VALUE, where the value can be JSON, and list them with hbench agents schema:

hbench agents schema mini-swe-agent
hbench run -d terminal-bench@2.0 -l 3 -a mini-swe-agent -m deepseek/deepseek-flash --ak max_tokens=4096

A run the site issues can't set options whose names end in _path, _file or _dir, because they would make the agent read files outside the task.

hbench installs most agents at a version it pins, so an agent's new release doesn't change a run; a version option installs another. Left blank, version installs the pinned version, whatever hbench's help says. The installers of antigravity-cli, cortex-code, cursor-cli, junie, mimo, muse-code and rovodev-cli take no version and always fetch the newest release.

Options a board holds#

The site checks two more things about an agent's options, on an issued run, an upload and Publish:

• No option may name a model or a model endpoint, such as claude-code's fallback_model. The site refuses one with model-option; Options that swap the model lists the names that count.

• On the named boards, claude-code always runs with disable_web_search: true, because its web search runs on Anthropic's side, out of the allowlist's reach. An issued run can leave the option out. The site refuses, with option-fixed, an issued run that sets it otherwise and a job published there that ran without it.

Models and keys#

Name the model with -m as provider/model, such as deepseek/deepseek-flash, anthropic/claude-opus-5-5 or openai/qwen2.5-coder:7b. The model decides the board, as in Name the model. mini-swe-agent refuses a model without its provider: Model name must be in the format provider/model_name.

The agent reads the provider's key from your shell:

Export a key from a file so it stays out of your shell history, such as export ANTHROPIC_API_KEY="$(cat ~/.anthropic-key)", and use a separate key with a spending limit. Keeping your model key safe explains where the key goes during a run.

Prompt templates#

Some agents take a prompt template: a Jinja template of up to 20,000 characters that wraps each task's instruction, with {{ instruction }} where the task's text goes. An issued run comes with its template as prompt.j2, and the site checks that every trial used it. Runs with different templates are different rows on a board. Only an issued run can carry a template or extra instructions: the site refuses to publish a job that used either.

Every agent also takes extra instructions: up to 8 passages, each up to 4,000 characters, added to every task. Use them for an agent without a template, such as terminus-2.

Run your own agent#

An agent that speaks the Agent Client Protocol (ACP) can go on a board too, from the ACP registry with -a acp:<id> or from a git repository with -a acp. Run your own harness or ACP agent shows how.

Built-in agents#

Reviewed on October 24, 2026.