Work with images
A user attaches a screenshot of an error and asks what went wrong. A tool returns a chart the agent should interpret. Both work in AXL: image content is passed to models that advertise multimodal support, and image results stream to the client alongside the rest of the run events.
To use images successfully:
- Select a model that supports image input.
- Upload the image through the attachment API or a supported channel.
- Include the attachment in the agent request.
- Ask a concrete question about it.
Attaching a picture does not describe a task. Tell the agent what to do with the image, like reading the error code in the red box.
The registry decides
AXL reads multimodal support from its model registry rather than probing the provider. When a registry entry does not advertise image input, AXL strips image content before sending and logs a warning. The run still succeeds.
A model that answers as though it never saw the picture usually means the registry row understates what the model can do, not that the upload failed. Nothing errors, the answer is just oddly generic.
It matters most for self-hosted models, where you write the registry row yourself. See local models.
Images and long sessions
Images occupy context. During compaction older images may be replaced with textual placeholders, so if a session will run long, ask the agent to record important conclusions from an image as text while it can still see the picture.
Next
- Manage long conversations - what compaction does to older content.
- Work with files - attachment storage and durable artifacts.
- - current attachment and run request schemas.