Agentic Engineering Workflow
This workflow enables you to transition from manual implementation to high-level system architecture by managing autonomous AI agents (like Devin) as "junior buddies." By shifting implementation to agents, you can scale a small team (e.g., 15 engineers) to handle the output of a much larger organization, aiming for 25% to 50% of pull requests to be AI-generated.
Core Principle: Bricklayer to Architect
Most engineering time is spent on "bricklaying": debugging Kubernetes errors, fixing port issues, or writing boilerplate code. Your goal is to move to "architecting": defining the problem precisely, mapping out the solution, and specifying trade-offs, while the agent handles the execution.
1. Task Delegation Framework
Do not hand agents "problems" (ambiguous high-level goals); hand them "tasks" (well-defined, verifiable units of work).
- Verifiability: Choose tasks that have an automated feedback loop (e.g., code that can be run, tests that can pass, or UI that can be previewed).
- The "Junior Buddy" Lens: Treat the agent like a talented but new junior engineer.
- Bad Prompt: "Fix our scaling issues." - Good Prompt: "I'm seeing a 404 error on the signup page. Research the logs in Datadog, reproduce the bug in a local environment, and suggest a fix."
2. Managing the Asynchronous "Fleet"
Do not watch the AI work action-by-action. To achieve massive productivity gains, you must manage multiple agents in parallel.
- The 5-Devin Rule: Aim to have up to 5 agents running at once.
- Morning Kickoff: Identify the 5 most discrete tickets in your sprint (e.g., Linear or Jira). Assign each to a separate agent session.
- Context Sharing: Use an integrated "Wiki" or index tool so the agent can learn the idiosyncrasies of your specific codebase (e.g., "how we handle multi-token prediction" or "our specific deployment operations").
3. The Integration Loop
Integrate the agent into your existing human workflows to maintain quality and oversight.
- Communication Channels: Interact via Slack for quick steering and GitHub for code review.
- The "Jagged Intelligence" Review: Be aware that AI has "jagged intelligence"—it may solve a complex algorithm but fail at a basic architectural convention.
- Review the Plan before execution. - Review the PR before merging.
- Interactive Planning: If an agent asks a question (e.g., "Should the button open in a new tab?"), answer immediately to keep the asynchronous momentum.
4. Onboarding and Documentation
Use agents to bridge the knowledge gap for human engineers.
- The Devin Wiki: Have the agent index the codebase and generate diagrams/explanations of complex modules (e.g., FP8 operations or networking abstractions).
- AI Mentorship: Use agents to answer "dumb questions" for new hires, such as "Where is the feature flag for the billing module located?"
Examples
Example 1: Bug Reproduction and Fix
- Context: A user reports that the sidebar links are broken on mobile.
- Input: Tag the agent on the Linear ticket with the specific error report.
- Application: The agent spins up a virtual machine, reproduces the mobile view, identifies the CSS conflict, and runs the linter.
- Output: A GitHub Pull Request with a screenshot of the fix in the mobile preview.
Example 2: Feature Implementation
- Context: You need to add a "Newsletter Feature" component to the web app.
- Input: "Modify the web app to feature this URL. Use the existing sidebar component. Make sure the link opens in a new tab."
- Application: The agent researches the sidebar code, creates a new component, and asks for clarification on styling. You provide 1-2 lines of feedback on the roundness of the button.
- Output: A ready-to-merge PR that matches the existing site architecture.
Common Pitfalls
- Watching the Pot Boil: Staying "synchronous" and watching the agent's terminal. This wastes your time. Kick off the task and come back when notified in Slack.
- Ambiguous Scoping: Giving a task that requires 50 architectural decisions without providing a starting point. Start with a "one-pointer" task to help the agent get familiar with the repo first.
- Ignoring the Trace: Not looking at the research steps the agent took. If an agent fails, check its "Playback" to see where its logic diverged from a human's.
- Over-Reliance on Base IQ: Assuming the AI knows your company's specific "messiness" (e.g., old COBOL or legacy wrappers). You must explicitly point it to the documentation for your "jagged" areas.