Research-Plan-Implement Loop

Created: 2026-06-02 09:10
#note

The Research → Plan → Implement loop (abbreviated RPI, a name coined by the HumanLayer project) is the foundational workflow pattern of Harness Engineering. Its governing principle, stated directly by Boris Tane, is that the agent must never write code until a written plan has been reviewed and approved. The pattern separates thinking from typing: architectural decisions are made and validated in a plan artifact before any implementation begins.

The Three Phases

Research directs the agent to read the relevant codebase deeply and to write its findings into a persistent markdown file rather than producing only a verbal summary in chat. This artifact becomes the human's review surface, allowing misunderstandings to be corrected before they propagate. The most expensive failure mode in AI-assisted coding is not bad syntax but an implementation that works in isolation while breaking the surrounding system; the research phase exists to prevent it.

Plan takes the research and produces a detailed specification — approach, code snippets, file paths to be modified, and trade-offs. The plan is then refined through the Plan Annotation Cycle until it is correct.

Implement executes the approved plan, ideally with test-driven development, marking tasks complete as it proceeds. Because every decision was made upfront, implementation becomes deliberately mechanical rather than creative.

Backflow Is Expected

The defining feature of RPI is that it is not waterfall. Backflow between phases is normal and expected: planning may reveal a design flaw and return to research; implementation may surface missing tasks and return to planning; verification may find bugs and return to implementation. This tolerance for iteration is what distinguishes RPI from rigid spec-driven development, which breaks the moment reality diverges from the specification.

graph LR
    R[Research] --> P[Plan]
    R -. "review & annotate" .-> H[Human]
    H -. "update spec" .-> R
    P --> I[Implement]
    I --> F[Finish / Verify]
    I -. "gaps found" .-> P
    P -. "design flaw" .-> R
    F -. "bugs found" .-> I

Convergent Evolution

A notable observation is that multiple practitioners arrived at this pattern independently. HumanLayer named it RPI; Jesse Vincent's Superpowers framework landed on the same shape; Martin Richards's Atelier is another instance; and Boris Tane's workflow follows the same structure. The convergence suggests the value lies in the discipline itself rather than in any single implementation — the harness just encodes that discipline so the engineer does not have to remember it each time.

References

1. Building Your Own Agent Harness — Martin C. Richards
2. How I Use Claude Code — Boris Tane
3. HumanLayer (GitHub)
4. Superpowers — Jesse Vincent (GitHub)

Tags

#harness_engineering #agentic_ai #rpi #spec_driven_development #ai_agents #workflow