Plan-Driven Agentic Engineering
Traditional AI coding is often "prompt-to-patch"—asking for a single fix. This skill shifts the workflow to a "teammate" model, where the human and agent collaborate on a structured plan before execution to ensure the agent can handle long-running, complex tasks without drifting or losing context.
The Workflow
1. Collaborative Planning (The plan.md Phase)
Before asking the agent to write a single line of production code, align on the logic.
- Create a
plan.md file: Write out the intended changes in markdown within your IDE.
- Context Loading: Explicitly provide the agent with relevant files, API documentation, or existing patterns (e.g., "Look at the iOS implementation to plan the Android port").
- Verifiable Steps: Break the plan into a checklist. Each step should have a clear "Definition of Done" that the agent can eventually verify itself.
- Iterate: Ask the agent, "Given this plan, what am I missing?" or "What are the biggest risks in this architecture?"
2. Execution Loop
Once the plan is finalized, move into execution.
- Delegation: Point the agent to the
plan.md and instruct it to execute a specific subset of tasks.
- Sandbox Execution: Allow the agent to run code, tests, and shell commands within a sandbox to validate its own work as it goes.
- Compaction: If the task is long-running (exceeding the context window), instruct the agent to "summarize the current state and remaining tasks into a fresh context" to maintain focus.
3. Mixed-Initiative Validation
The bottleneck of AI coding is often the human review. Shift the burden of proof to the agent.
- AI-Led Verification: Prompt the agent to "Write a test suite that proves your fix works" or "Verify your changes against these three edge cases."
- Visual Previews: For UI work, prioritize reviewing the image/preview output before diving into the code diff.
- Catching Configuration Errors: Use the agent to review infrastructure-as-code or configuration files specifically for "interesting mistakes" that humans often overlook in large PRs.
Strategic Applications
Vibe Coding (Rapid Prototyping)
Use the agent to build "throwaway" code to test an idea.
- The Pattern: Instead of writing a spec, ask the agent to "Build an interactive data viewer for this raw JSON" or "Vibe code a prototype of this animation in a standalone React file."
- Goal: Move from "talking about it" to "playing with it" in minutes.
Feature Porting
The most efficient way to scale across platforms (e.g., iOS to Android).
- The Input: Feed the agent the source code of the finished platform.
- The Prompt: "Analyze this iOS Swift implementation. Create a plan to port this to Kotlin, maintaining the same state management logic and UI flow."
Examples
Example 1: Solving a "Gnarly" Production Bug
- Context: A race condition in a high-traffic microservice that is difficult to replicate locally.
- Input: Provide the agent with the relevant service files, a recent stack trace, and Datadog logs.
- Application:
- Ask the agent to analyze the logs and propose 3 potential causes in
plan.md.
- Instruct the agent to write a stress-test script to attempt to reproduce the failure in the sandbox.
- Once reproduced, ask for a patch and a verification test.
- Output: A verified patch and a new regression test.
Example 2: The Sora App "Port" Pattern
- Context: Building an Android version of an existing, complex iOS app with only 2 engineers.
- Input: iOS repository and Android project scaffold.
- Application:
- Agent reads iOS source to understand business logic and API endpoints.
- Agent generates a 10-day roadmap of Kotlin tasks.
- Engineers use the agent to "translate" views while they focus on platform-specific performance tuning.
- Output: A GA-ready Android app in under 30 days.
Common Pitfalls
- The "Fresh Computer" Trap: Giving an agent a task without the necessary passwords, API keys, or environment dependencies. Treat the agent like an intern: if they don't have the environment set up, they can't work.
- Ignoring the Review Bottleneck: Spending 5 minutes writing code with AI but 60 minutes reviewing it. Instead, instruct the agent to "Explain this diff to me like I'm a senior reviewer and highlight potential side effects."
- Over-relying on Prompting: If a task fails, don't just re-prompt. Check if the agent needs a better "harness" (e.g., a better test runner or access to a specific documentation site).
- Dumb Tasks Only: Using AI for only "easy" tasks. The highest value is found in giving the agent your "hardest problems" and collaborating on the plan.
1---2name: plan-driven-agentic-engineering3description: A workflow for collaborating with AI coding agents (like Codex) on complex, multi-step software tasks. Use this when porting features between platforms, building rapid prototypes ("vibe coding"), or resolving "gnarly" production bugs that require deep context.4---56# Plan-Driven Agentic Engineering78Traditional AI coding is often "prompt-to-patch"—asking for a single fix. This skill shifts the workflow to a "teammate" model, where the human and agent collaborate on a structured plan before execution to ensure the agent can handle long-running, complex tasks without drifting or losing context.910## The Workflow1112### 1. Collaborative Planning (The plan.md Phase)13Before asking the agent to write a single line of production code, align on the logic.14- **Create a `plan.md` file:** Write out the intended changes in markdown within your IDE.15- **Context Loading:** Explicitly provide the agent with relevant files, API documentation, or existing patterns (e.g., "Look at the iOS implementation to plan the Android port").16- **Verifiable Steps:** Break the plan into a checklist. Each step should have a clear "Definition of Done" that the agent can eventually verify itself.17- **Iterate:** Ask the agent, "Given this plan, what am I missing?" or "What are the biggest risks in this architecture?"1819### 2. Execution Loop20Once the plan is finalized, move into execution.21- **Delegation:** Point the agent to the `plan.md` and instruct it to execute a specific subset of tasks.22- **Sandbox Execution:** Allow the agent to run code, tests, and shell commands within a sandbox to validate its own work as it goes.23- **Compaction:** If the task is long-running (exceeding the context window), instruct the agent to "summarize the current state and remaining tasks into a fresh context" to maintain focus.2425### 3. Mixed-Initiative Validation26The bottleneck of AI coding is often the human review. Shift the burden of proof to the agent.27- **AI-Led Verification:** Prompt the agent to "Write a test suite that proves your fix works" or "Verify your changes against these three edge cases."28- **Visual Previews:** For UI work, prioritize reviewing the image/preview output before diving into the code diff.29- **Catching Configuration Errors:** Use the agent to review infrastructure-as-code or configuration files specifically for "interesting mistakes" that humans often overlook in large PRs.3031## Strategic Applications3233### Vibe Coding (Rapid Prototyping)34Use the agent to build "throwaway" code to test an idea.35- **The Pattern:** Instead of writing a spec, ask the agent to "Build an interactive data viewer for this raw JSON" or "Vibe code a prototype of this animation in a standalone React file."36- **Goal:** Move from "talking about it" to "playing with it" in minutes.3738### Feature Porting39The most efficient way to scale across platforms (e.g., iOS to Android).40- **The Input:** Feed the agent the source code of the finished platform.41- **The Prompt:** "Analyze this iOS Swift implementation. Create a plan to port this to Kotlin, maintaining the same state management logic and UI flow."4243## Examples4445**Example 1: Solving a "Gnarly" Production Bug**46- **Context:** A race condition in a high-traffic microservice that is difficult to replicate locally.47- **Input:** Provide the agent with the relevant service files, a recent stack trace, and Datadog logs.48- **Application:** 49 1. Ask the agent to analyze the logs and propose 3 potential causes in `plan.md`.50 2. Instruct the agent to write a stress-test script to attempt to reproduce the failure in the sandbox.51 3. Once reproduced, ask for a patch and a verification test.52- **Output:** A verified patch and a new regression test.5354**Example 2: The Sora App "Port" Pattern**55- **Context:** Building an Android version of an existing, complex iOS app with only 2 engineers.56- **Input:** iOS repository and Android project scaffold.57- **Application:**58 1. Agent reads iOS source to understand business logic and API endpoints.59 2. Agent generates a 10-day roadmap of Kotlin tasks.60 3. Engineers use the agent to "translate" views while they focus on platform-specific performance tuning.61- **Output:** A GA-ready Android app in under 30 days.6263## Common Pitfalls64- **The "Fresh Computer" Trap:** Giving an agent a task without the necessary passwords, API keys, or environment dependencies. Treat the agent like an intern: if they don't have the environment set up, they can't work.65- **Ignoring the Review Bottleneck:** Spending 5 minutes writing code with AI but 60 minutes reviewing it. Instead, instruct the agent to "Explain this diff to me like I'm a senior reviewer and highlight potential side effects."66- **Over-relying on Prompting:** If a task fails, don't just re-prompt. Check if the agent needs a better "harness" (e.g., a better test runner or access to a specific documentation site).67- **Dumb Tasks Only:** Using AI for only "easy" tasks. The highest value is found in giving the agent your "hardest problems" and collaborating on the plan.
samarv