An Effective AI Agent Pattern: Plan, Tasks, Implementation

For source files view the repository: https://github.com/gbro3n/agent-flow

The agent-flow repository serves as a template for a flow I've recently been using for agent-assisted coding - with impressive results.

Even with the best models, agent-assisted coding can be hit and miss. A good AGENTS.md that covers most of the mistakes you find agents making can go a long way, but this flow has been producing outsized results over the last couple of weeks.

It takes a little more discipline than hacking away in the chat window until you get what you want, but this process makes a "one shot" success more likely, and really pays off when things get complicated and the agent needs more of a steer.

The result is that I've been moving through the build of features of a size and complexity that I'd normally expect agents to struggle with, with incredible speed and precision. I'm confident that with this flow, agentic coding provides a huge net positive in my productivity, where before I could not have said with certainty that it made me faster over the long term.

I work mostly with Claude Sonnet 4.6 and Opus 4.6, but I started this flow after initial testing with these newer Claude versions, so I am confident that the flow has made a difference. This process gives enough structure to the agent that I'm less likely to feel the need to reach for Opus, and that I can complete tasks faster and at less cost with Sonnet.

Getting Started

• Copy the contents of this repository into your project root.
• Read AGENTS.md — it's the core document that drives the agent's behaviour.
• Start a planning session by typing plan in the chat window.

The Process

For full detail read AGENTS.md — it's not that long. Here's a summary:

AGENTS.md is set up to make the process clear to your agent:

memory.md

A persistent space for the agent to record anything useful for the long-term life of the project. Unlike plan.md and todo.md, memory.md is not reset between tasks — think of it as the agent's long-term project memory.

plan.md

Start here. Write out your requirements in plan.md before involving the agent.

The benefits are two-fold:

• You are more likely to produce clear and detailed requirements when writing in a file than when typing in the chat window.
• Your input persists beyond the agent's context window and is easy to refer back to for both you and the agent.

plan.md uses a conversational format with a summary section at the top. Both you and the agent write responses in this file. You can also add [comment: <text>] inline in the agent's responses to steer it.

[user]: <instruction for agent>

[agent]: <agent response>

[user]: <instruction for agent>

[agent]: <agent response>

AGENTS.md instructs the agent not to implement anything until it has produced a task list and received your explicit confirmation.

To start a planning session, type plan in the chat window.

Tip: Reset plan.md and the agent context after completing a feature.

todo.md

Once planning is complete, the agent produces a task list in todo.md. This gives the agent a concise, concrete set of anchors to work from as it moves through the implementation.

To start implementation, type implement in the chat window.

Tip: Reset todo.md after completing a group of related features.

Conclusion

I would encourage you to give a flow like this a try if you haven't already and hopefully you'll see improved results in your agentic workflows.