Orchestrating Development: How AI Agents are Changing Large-Scale Workflows
The role of a developer is rapidly shifting. We are moving from "writing all the code" to "orchestrating the development process." In large-scale, complex environments like GitLab, this change is not just about speed—it's about consistency, precision, and handling complexity through AI-human collaboration.
Here is how we are managing that shift right now.
The Shift: From Doing to Orchestrating
Instead of manual tasks, we use a structured agent-driven workflow. The goal is to define the intent and let AI handle the heavy lifting, while human experts maintain oversight.
1. Guidelines as Code
We don't rely on tribal knowledge. Rules of the road are documented directly in the repository via CLAUDE.md and AGENTS.md. These files tell the AI exactly how to:
- Handle security fixes safely.
- Manage CI/CD failures without manual intervention.
- Apply proper labeling and commentary standards.
By treating these guidelines as code, every agent we spawn immediately understands our standards, reducing the need for back-and-forth corrections.
2. Specialized Agents for Specialized Tasks
We use specific agents for distinct parts of the lifecycle:
- Plan Agent: For architecting changes and considering trade-offs before a single line of code is written.
- General Purpose Agent: For complex, multi-step research or code discovery.
- Code Review Agent: To perform high-effort correctness and efficiency reviews based on established project principles.
3. The Human-in-the-Loop
This is the most critical part. The AI is a powerful assistant, not a replacement. Our workflow relies on:
- Clear Intent: The human defines the "what" and "why."
- Agent Execution: The agent produces the draft, applies fixes, or runs the search.
- Human Verification: The human reviews the output, verifies against project standards, and orchestrates the next step.
The Toolkit for Modern Development
To work effectively in this environment, you need more than just coding skills. You need a mix of:
- Project Domain Knowledge: Deep familiarity with your codebase and project-specific guidelines.
- Git CLI Mastery: Navigating branches, rebasing, and managing pull requests efficiently from the command line.
- AI Orchestration: Knowing which agent to use, when to use it, and how to prompt it for the best result.
- CI/CD Debugging: The ability to triage and diagnose pipeline failures—understanding what is infrastructure, what is flaky, and what requires a code fix.
Final Thoughts
This isn't about automating away the developer; it’s about elevating the developer's role from a low-level implementer to an architect and conductor. By leveraging AI to handle repetitive tasks and complex searches, we free ourselves to focus on the high-level design and critical architectural decisions that truly move the needle.
Top comments (0)