Skip to main content

Research → Design → Plan → Implement

February 17, 2026

1 min read

workflowagentsharnessrecipe

Recipe

For features that take 4–6 hours, this simple sequence works well:

  1. Research: gather constraints, edge cases, and existing patterns.
  2. Design: decide the shape of the solution (interfaces, data flow, tradeoffs).
  3. Plan: break the work into small, ordered steps with clear checkpoints.
  4. Implement: execute one step at a time, validating as you go.

Why this helps agent workflows:

  • Context window control: each phase keeps only relevant information in focus.
  • Quality: decisions are made before coding, so fewer surprises appear late.
  • Throughput: smaller checkpoints reduce rework and make progress visible.

A useful rule of thumb:

If a feature is too big to explain in one short plan, it is too big to implement in one pass.

Split it, then run the same sequence again.

Read next

The Gradient

The feedback loop that makes agentic coding work is just gradient descent in disguise.

Put this into practice: spokA lightweight spec-driven workflow for AI coding agents: explore an idea, propose a chunked change, apply one slice, then fold the result back into living specs.

Share: