Skip to content

Goal-Driven Execution principle

The Goal-Driven Execution principle is a guideline for AI coding agents (and engineers) that prioritizes achieving verifiable results over simply producing code output^[001-TODO__Andrej_Karpathy_Skills_-AI_Coding_Agent_行为框架.md]. It counteracts the tendency of agents to engage in "hope-based" programming, where a task is declared finished once the code is written, without confirming that the code actually solves the problem^[001-TODO__Andrej_Karpathy_Skills-_AI_Coding_Agent_行为框架.md].

Core Concept

The core philosophy of Goal-Driven Execution is that an action is only complete when the specific objective is met and verified^[001-TODO__Andrej_Karpathy_Skills_-AI_Coding_Agent_行为框架.md]. This principle dictates that the workflow must transition from open-ended generation to a closed-loop cycle involving reproduction, repair, and verification^[001-TODO__Andrej_Karpathy_Skills-_AI_Coding_Agent_行为框架.md].

Instead of responding with "I implemented it" immediately after generating a solution, the agent (or developer) must operate based on explicit success criteria^[001-TODO__Andrej_Karpathy_Skills_-_AI_Coding_Agent_行为框架.md].

Execution Workflow

Goal-Driven Execution typically follows a strict four-step loop^[001-TODO__Andrej_Karpathy_Skills_-_AI_Coding_Agent_行为框架.md]:

  1. Reproduce: First, identify and understand the issue or requirement in the current context.
  2. Fix: Apply the specific changes necessary to address the issue.
  3. Verify: actively test the changes to confirm the issue is resolved.
  4. Stop: Conclude the task only after verification is successful.

Comparison with Unstructured Execution

Without the Goal-Driven principle, agents or developers often fall into a "generate and pray" pattern. The following contrasts the two approaches^[001-TODO__Andrej_Karpathy_Skills_-_AI_Coding_Agent_行为框架.md]:

Unstructured ("Hope it works") Goal-Driven Execution
Mindset: Task is finished when code is written. Mindset: Task is finished when the result is verified.
Result: "I implemented it." Result: Verified confirmation of the fix.
Reliability: High risk of introducing bugs or missing the actual requirement. Reliability: Higher confidence that the specific problem is solved.
  • [[Andrej Karpathy Skills - AI Coding Agent 行为框架]]
  • [[Surgical Changes]]
  • [[Think Before Coding]]

Sources

  • 001-TODO__Andrej_Karpathy_Skills_-_AI_Coding_Agent_行为框架.md