AI SDK adds HarnessAgent for Pi, Claude, Codex, and OpenCode

• cramforce's launch note says AI SDK's new HarnessAgent system gives one interface for Pi, Claude, Codex, OpenCode, and other coding harnesses.
• cramforce's launch note says HarnessAgent uses official SDKs under the hood, and cramforce's Codex reply says an existing Codex subscription can pay for software-factory token usage.
• cramforce's deployment reply says the system is for local and cloud runs, while lgrammel's sandbox example shows a local Docker sandbox on Mac using Apple Container.
• lgrammel's factory post says work has started on a software factory for AI SDK, and lgrammel's clarification says that means a factory that builds parts of AI SDK itself.
• lgrammel's sandbox interface reply says Experimental_SandboxSession and HarnessSandboxV1 are public interfaces, with Modal, E2B, and Daytona named as possible community sandbox packages.

Christmas came early for coding-agent harness nerds: Vercel's June changelog describes HarnessAgent as a single API for established agent harnesses, the adapter docs already list Claude Code, Codex, Deep Agents, OpenCode, and Pi, and the UI docs say harness sessions own their own conversation state instead of replaying chat history into a model. The buried gotcha is in the sandbox reference: passing experimental_sandbox gives tools an execution environment, but tool code still runs in the app process unless it explicitly delegates work into the sandbox.

Vercel's HarnessAgent changelog says AI SDK 7 adds a single API for running established harnesses, with switching framed as the harness-level version of swapping model providers.

The AI SDK Harnesses overview defines a harness as a complete agent runtime, not a model call. The runtime owns workspace access, built-in coding tools, native session state, compaction, permission flows, and runtime-specific configuration.

The docs split the system into four pieces:

• HarnessAgent: the application-facing AI SDK agent implementation.
• Harness adapter: the connector for Claude Code, Codex, Pi, OpenCode, or another runtime.
• Sandbox provider: the isolated filesystem and process environment.
• Session: the live state for conversation history, workspace, and approvals.

The current adapter docs list five shipped adapters and three coming soon:

• Claude Code, @ai-sdk/harness-claude-code, sandbox bridge.
• Codex, @ai-sdk/harness-codex, sandbox bridge.
• Deep Agents, @ai-sdk/harness-deepagents, sandbox bridge.
• OpenCode, @ai-sdk/harness-opencode, sandbox bridge.
• Pi, @ai-sdk/harness-pi, host process.
• Coming soon: Amp, Goose, Mastra.

The Claude Code adapter docs say it connects through @anthropic-ai/claude-agent-sdk; the Codex adapter docs say it connects through @openai/codex-sdk; the OpenCode adapter docs say it starts an OpenCode server inside the sandbox and streams session events back over WebSocket.

The token-billing detail is unusually concrete for a harness abstraction. cramforce's Codex reply says a Codex subscription can pay for the tokens a software factory uses.

That matches the shape of the implementation in the docs: the Codex adapter runs through the Codex SDK, and the Claude Code adapter runs through Anthropic's Claude Agent SDK. cramforce's launch note says the official SDK path is why subscriptions "just work."

cramforce's deployment reply says HarnessAgent is not Vercel-only and can run local or cloud, though he adds that software factories run better in the cloud.

The local path is already visible. lgrammel's local sandbox example uses createAppleContainerSandbox, experimental_sandbox, and a shell tool to create and run a Node.js file inside a Mac local Docker-style sandbox.

The extension seam is public. lgrammel's sandbox interface reply names Experimental_SandboxSession and HarnessSandboxV1 as public interfaces and calls out Modal, E2B, and Daytona as possible community sandbox packages.

One composition pattern is already possible: cramforce's tool-wrapper reply says a developer can make a tool that uses HarnessAgent directly when the harnesses have isolated work.

A separate cramforce progress reply says related work is in progress, with no promise attached.

pauliusztin_ framed the portability boundary above the harness. pauliusztin_'s context-layer post argues that models are commoditizing fast, harnesses already have, and the thing to own is the context layer.

His proposed context layer contains:

• unified memory, business logic, serving layer, and a disposable harness;
• MCP servers or skills for tools, resources, prompts, and domain knowledge;
• a starting point as simple as markdown files;
• a later path to knowledge graphs, filesystem storage, and semantic plus BM25 retrieval;
• one store, with MongoDB given as the example.

zeeg's counterweight is that harnesses are not automatically better than native providers for every agent job. zeeg's harness caveat says the important part is why a model performs differently in a specific scenario, not a table of scores.

zeeg's table reply adds that the result was hard to compress into a table. pauliusztin_'s testing quip gives the same failure mode a shorter name: the AI version of "works on my machine" is testing a few prompts and deciding it looked good.