The Agent Harness Is Now a Product. Should Indie Devs Rent One?

The harness is now a rented product, and for most small teams that is good news. Microsoft made the Copilot harness available through Microsoft Foundry on 28 October (Microsoft), and Cursor shipped its Agent SDK on 20 October (Cursor). The question is not whether these harnesses are good. It is whether renting one leaves you with a product or with a dependency.

TL;DR

• Microsoft opened the Copilot harness through Foundry on 28 October; Cursor shipped its Agent SDK on 20 October.
• The model is becoming interchangeable, while the harness holds instructions, tools, sandbox, evals and routing.
• Renting buys months of work and hands another company control of your agent loop.
• Keep prompts, tools, eval sets and data portable so the decision stays reversible.
• Event-driven work, such as PR review or ticket triage, fits small teams particularly well.

Disclosure: ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 29 Oct 2026; ZeroRaised has no commercial relationship with the companies named.

Agent harness is the layer around a model that supplies the instructions, tools, sandbox, evaluation loop and model routing which turn a raw model into a working agent.

What did Microsoft and Cursor actually ship?#

Microsoft's harness is not new code built for a launch. It already powers Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI and VS Code, and it was built by GitHub engineers (Microsoft). What changed on 28 October is that outside developers can now use it through Foundry.

The pitch is multi-model. Microsoft says the harness supports more than 20 frontier models plus bring your own key, and that it routes work to models (Microsoft). Microsoft's own definition is the cleanest sentence in the announcement: "If the model supplies the intelligence, the harness is everything around the model" (Microsoft).

Cursor's Agent SDK is a different shape. It is "the framework we use to build our own managed agents like Bugbot, Security Reviewer, and Rollouts" (Cursor). Agents are defined in code: prompts in Markdown, tools as TypeScript functions, versioned and reviewed like any other code (Cursor). It responds to pull request events from GitHub, GitLab or Bitbucket, supports evals and event replay, and supports subagents (Cursor). You can deploy with agent-sdk deploy to Cursor infrastructure or run it on your own as a single Node process in a container, and it is generally available (Cursor).

Why is the harness the valuable part?#

Because the model is the easy thing to swap. Microsoft says plainly that "a frontier model today might fall behind tomorrow" (Microsoft). If that is true for Microsoft, it is certainly true for you.

The durable work sits in the harness: which tools exist, what the instructions say, how the sandbox is bounded, how output is evaluated, and when a cheaper model handles the boring step. That is the same argument behind AI agent vs AI workflow: the useful question is not which model, but what the system is allowed to do.

Cursor states the ownership boundary well: "Your instructions, tools, and evals are files you review, test, and version like the rest of your code. You write the instructions and tools, and Cursor runs the model loop that calls them" (Cursor docs). That sentence is also the portability clause. If those files are ordinary Markdown and TypeScript, moving them later is tedious rather than impossible.

What do you get by renting?#

Time, mostly. Cursor says a team built kb-reviewer, an agent that checks every pull request against a knowledge base of lessons, "in about a day" (Cursor). A solo founder would spend far longer than a day on sandboxing, event handling, retries, evals and deployment.

You also get the operational boring parts. Microsoft runs the agent loop and sandbox, and says you bring the instructions, skills, tools and model choice (Microsoft). For a two-person team, that is a real trade.

And you get standards. Microsoft says "a harness worth its salt has support for open standards like MCP and agent skills", and that the runtime was recently ported to Rust (Microsoft). MCP support matters because it means your tool layer can look like the tool layer everyone else writes. If you have not built one before, start with how to build an MCP server.

Where does renting go wrong?#

Three places.

First, pricing and packaging. A rented loop turns a fixed engineering cost into a recurring one, and the vendor can change the meter. That is not a reason to avoid it, but it is a reason to model the cost per task before you design a business around it, and to read how to reduce LLM costs before volume arrives.

Second, lock-in through the parts that are not portable. Prompts and tools may be files, but the eval harness, event replay, deployment target and observability often are not. Ask what you would keep if you left, and test that answer before you ship.

Third, the product shape. If your only claim is "an agent for X" when anyone can rent the same harness and point it at the same model, your moat is thin. That is a positioning problem, and it is the same one covered in how to build a defensible AI product.

What this means for bootstrapped founders#

Rent first, but rent deliberately.

• Pick an event-driven job. Agents triggered by a pull request, a ticket, a new file or a failed check have a clear start, a clear end and measurable success. Both launches point this way: Cursor ships PR events, and Bugbot is the example it names.
• Treat prompts, tools and evals as your source of truth. Keep them in your repo, reviewed like code. That is how you stay portable even while renting the loop.
• Sandbox before you scale. Renting does not remove the need to bound what the agent can do. The agent sandbox guide covers the minimum.
• Review generated work like a teammate's. The AI code review checklist is the habit that keeps quality from drifting.
• Price the harness into your plan. If you sell the output, know your cost per task at your expected volume, not at demo volume.

The blunt version: use Foundry or Cursor to get to a working agent in weeks, then decide whether the loop is worth owning. Do not confuse a fast start with a moat.

Where ZeroRaised fits#

ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you have shipped an agent or a tool that plugs into one of these harnesses, you can submit your product and describe what it actually does. The pricing page explains the current options, and the newsletter is where the next batch of analysis lands.

No promises of traffic. A launch page is a place to state the facts clearly: which harness you use, which models you support, what your agent is allowed to touch and what it costs.

Sources#

• Microsoft: Introducing the Microsoft Copilot harness
• Cursor: Introducing the Agent SDK
• Cursor: Agent SDK documentation