image_generation

Image Generation Agents

ActiveHarness agents can generate images natively — no custom_llm_backend/RubyLLM required. Enable it with image true and use the same model/fallback DSL as any other agent.

class ImageAgent < ActiveHarness::Agent
  image true
  size  "1024x1024"

  model do
    use      provider: :openrouter, model: "openai/gpt-5-image-mini"
    fallback provider: :openai,     model: "gpt-image-1"
  end
end

result = ImageAgent.call(input: "A watercolor fox in a snowy forest")

result.output      # => base64 string, "data:image/png;base64,..." or an HTTPS URL (provider-dependent)
result.processed   # => same as output (default format is :text)

Image APIs take a single prompt string — there's no separate system/user message split like in chat. So if the agent declares a system_prompt, it's prepended to @input (joined by a blank line) to form the final prompt; without one, @input is sent as-is.

class ImageAgent < ActiveHarness::Agent
  image true

  # Base style guidance — applies to every request.
  system_prompt "Watercolor style, soft pastel colors, no text overlays."

  model do
    use provider: :openrouter, model: "openai/gpt-5-image-mini"
  end
end

ImageAgent.call(input: "A fox sitting in a snowy forest")
# final prompt sent to the provider:
#   "Watercolor style, soft pastel colors, no text overlays.
#
#   A fox sitting in a snowy forest"

system_prompt accepts a plain string, a Prompt class, or a lambda — same as text agents (see Prompts) — so the base style can itself be dynamic (e.g. built from @context/@params).

Supported Providers

Only two providers currently support image generation:

Provider	Class	Example models
:openai	Providers::Images::OpenAI	"dall-e-2", "dall-e-3", "gpt-image-1"
:openrouter	Providers::Images::OpenRouter	any image-capable OpenRouter model, e.g. "openai/gpt-5-image-mini", "google/gemini-2.5-flash-image"