The Agentic Workflow

Unlike autocomplete tools (like GitHub Copilot) that predict the next few lines of text, Dodo behaves like a semi-autonomous developer. It uses a structured loop of reasoning to solve large-scale problems.

The Loop: Think → Plan → Act

When you give Dodo a task, it doesn't just start writing code. It follows a rigorous mental process:

graph TD
    User[User Request] --> Think[Thinking & Analysis]
    Think --> Plan[Create/Update Plan]
    Plan --> Tool[Execute Tool]
    Tool --> Result[Tool Result]
    Result --> Think
    Think --> Done[Task Completion]

1. Thinking (Reasoning)

Before touching any file, the agent analyzes the request. It asks: - "Do I have enough context?" - "What files are relevant?" - "Is this risky?"

This internal monologue is streamed to you in real-time, so you know exactly why the agent is making a decision.

2. Planning

For complex tasks (like "Refactor this module"), Dodo creates a structured plan. - It breaks the goal into step-by-step milestones. - It won't proceed to step 2 until step 1 is verified. - You are involved: You can see the plan and (in future versions) edit it before execution.

3. Tool Execution

Dodo interacts with your computer using Tools. It cannot do anything "magic" outside of these defined capabilities: - read_file: Read file contents. - codebase_search: Find code semantic search. - run_cmd: Execute terminal commands (tests, builds). - replace_file_content: Edit code with surgical precision.

4. Verification

Crucially, Dodo checks its work. After editing a file, it might run the build command or the unit tests to ensure it didn't break anything. If verification fails, it self-corrects and tries again.

Why this matters?

This workflow allows Dodo to handle tasks that require long-term memory and multi-step reasoning, such as: - "Upgrade this dependency and fix all breaking changes." - "Implement this feature across the backend and frontend." - "Debug this error by adding logs, running the app, and reading the output."