HarnessAgent

After the refactor, the built-in file system and shell operations are still isolated and executed inside the sandbox.

The fx adapter of HarnessAgent in AI SDK underwent an architecture-level refactor: harness and model requests no longer run in the sandbox but run in the same process as the application.

The new adapter supports generating responses that conform to a specified schema (structured output).

The new adapter supports sending new instructions to adjust the agent's direction while a turn is still running (mid-turn intervention).

The new adapter supports specifying which built-in tools a proxy is allowed to use (built-in tool filtering).

The official documentation currently focuses on integration methods and has not yet provided explanations of the boundaries of each adapter's image understanding capabilities or performance comparisons.

For projects already using HarnessAgent, the main cost of migrating to image input is rewriting existing plain-text messages into the content array form.

It is recommended to run small-sample comparisons separately for target scenarios (screenshot-to-code, error diagnosis, architecture diagram understanding) and to watch for whether each harness has implicit limits on image size, format, or count.

The unified interface only guarantees that messages reach the underlying vision model; it does not guarantee that different agents produce consistent output quality for the same image.

All harness adapters in the AI SDK's HarnessAgent layer can now accept image input.

HarnessAgent image input uses the standard AI SDK message format, and the content array can contain both { type: 'text' } and { type: 'file' } content blocks.

Image data is handled by the vision model underlying each harness rather than passing through the sandbox; the sandbox only hosts the agent's file system and command execution environment.

agent.getSandboxTemplate() returns the sandbox image template required by that harness, ensuring the runtime has the toolchain the agent depends on.

Java teams choosing AgentScope should consider three points: whether they already have Spring AI infrastructure, whether the task requires complex tool orchestration, and whether a human confirmation step must be introduced; if the latter two hold, the combination of ReActAgent and HarnessAgent is worth evaluating.

HarnessAgent consolidates engineering practices that were previously scattered across the application layer (such as file-purpose conventions and channel abstraction) into built-in framework capabilities, reducing the cost for teams to build their own runtime environment.

HarnessAgent provides a channel mechanism responsible for message flow between the agent and external systems, and offers context management strategies to address information trade-offs in long tasks.

HarnessAgent introduces the concept of a workspace, defining the boundary of files and resources an agent can operate on, with clear conventions for the purpose and storage structure of each file in the workspace, making agent behavior auditable and reproducible.

This version adds the ability to automatically persist validated HarnessAgent outputs to workflow results.