Research-Plan-Implement Pipeline (RPI Pipeline)

The RPI pipeline is a working method that requires an AI agent to go through three steps in order—research, plan, and implement—before starting a task. Think of a carpenter: instead of hammering in nails right away, they first walk the site (research), draw up blueprints (plan), and only then build the frame (implement).

For an agent working with code, the research step means opening the repository first to check what language it uses, what conventions existing tests follow, and what the actual build and run commands are. Next, the agent plans what to build and how, then writes the code and runs it to verify the results. This order lets the agent add new code in exactly the right place and format without touching production code.

This sequencing proves especially valuable when instructions are vague. Even if a person simply asks the agent to "write tests" without much detail, the agent reads and understands the repository on its own before acting, which reduces the chance of off-target results.

Microsoft said its open-sourced unit-test generation agent goes through planning, writing, execution, and verification via the Research-Plan-Implement (RPI) pipeline. It handled 79 out of 89 (88.8%) ambiguous prompts, a significant improvement in failure rate over a conventional agent that writes code immediately (66.3%). A common misconception is that RPI is not a proprietary technology belonging to any one company, but rather a general concept referring to the workflow of researching before acting.

When assigning a task to a coding agent, try breaking your request into these steps.

Step 1 (Research): First check and summarize the language this repository uses, its test framework, and the naming conventions of existing test files. Don't write any code yet.

Step 2 (Plan): Based on what you just found, plan which tests to add to which files.

Step 3 (Implement): Now write the test code according to the plan, run it, and report the results.

You'll notice that splitting the task into these stages, rather than asking for everything at once, reduces off-target results.