What happened
Vercel says the AI SDK HarnessAgent layer now accepts image inputs across every harness adapter it makes available. The change is at the agent-layer interface rather than a newly announced vision model: an application sends an image alongside text, and the vision-capable model underneath the chosen harness processes it.
The input contract is the standard AI SDK message format. A user message contains a content array whose parts can mix text and files. In Vercel’s example, the text part asks the agent to describe an image in one sentence. The following file part declares type: 'file', identifies the MIME type as image/png, and supplies bytes read from a local image.png file. That mixed message is passed to agent.generate together with an existing session.
The example constructs a HarnessAgent with the Codex harness. It then creates a Vercel network sandbox session using Node 24, exposes port 4000, obtains a sandbox template from the agent, creates an agent session, and calls generation. The point of the example is not a new sandbox feature; it demonstrates that image data can travel through the same harness-agent session flow as text.
Vercel lists Claude Code, Cline, Codex, Cursor, Deep Agents, fx, GitHub Copilot, Grok Build, OpenCode, and Pi as supported adapters. For an application that can select among those harnesses, image input no longer requires a different top-level message shape for each listed option. The changelog directs developers to the HarnessAgent image-input documentation to begin.
The announcement does not specify a version number, pricing, file-size limits, image-resolution limits, model-specific capabilities, or performance characteristics. It also does not claim that each underlying model will interpret a given image equally well. Its stated change is narrower: the shared HarnessAgent interface now accepts images for all listed adapters.
What the archive adds
This is a new feature on a known Vercel agent-runtime thread. The earlier Field Note, See every agent session in AI Gateway logs., described Vercel moving inspection from isolated requests to multi-turn sessions. This update adds another kind of state—an image—to a HarnessAgent session input, rather than adding new session observability.
Trend signal
The release supports Harness architecture outlives model choice. Vercel is placing a stable message contract above multiple agent harnesses and their underlying models: the application submits text and a file once, while adapter and model selection sit below that boundary. The evidence is limited to image input, but it is a concrete example of product value accruing at the harness interface rather than in any one model identity.
Why it matters
Visual context is common in software work: an agent may need to inspect a screenshot of a bug, a design mockup, a diagram, or a rendered page. Before a shared interface, a team supporting multiple coding-agent integrations can end up treating visual attachments as adapter-specific plumbing. This update makes the input shape portable across Vercel’s listed HarnessAgent adapters.
That portability is useful when a team wants to trial or switch harnesses without rewriting the part of its workflow that packages multimodal tasks. It also creates a clearer separation of concerns: the application owns the message and session flow; the selected harness and its underlying vision model handle interpretation. Builders should still evaluate the result at the adapter-model combination they intend to run, because a common transport contract is not evidence of common visual reliability.
What happens next
- More HarnessAgent examples will use mixed visual and textual task state. (3 months, 65%). Watch for: additional documented workflows beyond image description.
- Teams will compare adapters on vision-task outcomes, not only text-agent behavior. (6 months, 60%). Watch for: adapter-specific evaluation guidance or capability notes.
Limits and uncertainty
The public changelog establishes that Vercel accepts image inputs through the listed adapters, but it does not establish which underlying models are used, their image constraints, or whether every adapter has identical support for formats and session configurations. It does not report latency, cost, accuracy, or reliability measurements. The sandbox code is an example using Codex and a Vercel network sandbox, not proof that a sandbox is required for image input or that all adapters behave identically. Teams should verify supported media types, limits, and model behavior in the documentation and their own environment.
Skill to build now
Create one small multimodal acceptance test for an agent workflow this week. Send the same screenshot plus a bounded text request—such as identifying a visible UI regression or describing a chart—using the mixed content array. Run it through the harnesses you are considering, record whether the agent identifies the relevant detail, and retain the image and expected answer as a regression fixture. That turns a newly portable transport layer into a testable product decision.
Reader takeaway
HarnessAgent now gives Vercel AI SDK users one way to send images and text to every listed harness adapter. This removes a layer of multimodal integration work, but not the need to test the actual vision model and harness chosen for a production workflow.