In recent years, generative artificial intelligence has transformed how we develop software. Autonomous agents capable of writing code, running tests, and deploying changes have shown enormous potential. However, coordinating multiple agents remains a technical and operational challenge. Most current solutions rely on centralized orchestrators — Temporal, Airflow, RabbitMQ — adding complexity, cost, and single points of failure. Yet there is a more elegant and minimalist alternative: using labels in your issue tracker as a distributed state machine.
The idea is as simple as it is powerful: each task is represented by an issue in GitHub, GitLab or Jira. The state of that task is encoded in a label: ai-backend, ai-review, ai-qe. Each AI agent watches for issues bearing its label. When it finds one, it grabs the context, performs its work, and upon completion changes the label to indicate the next step. There are no direct agent-to-agent calls, no message queues, no external databases. The issue tracker itself acts as queue, audit log, and synchronization mechanism.
This approach aligns perfectly with the philosophy of custom software development that we promote at Q2BSTUDIO. When we build software solutions for our clients, we look for architectures that are simple, maintainable, and leverage existing infrastructure. A label-based pipeline meets those requirements: no new platforms to learn, developers already know the issue tracker, and the entire history is naturally documented in each issue's comments.
The lifecycle of an agent is an infinite loop of six phases: watch, grab, think, work, check, and hand off. Each agent is configured via a system prompt that defines its identity, skills, boundaries, and transition rules. For example, a backend agent knows it must never touch frontend files; a review agent knows it should not write implementation code. These clear boundaries prevent context drift and keep each agent within its area of responsibility.
One of the most attractive aspects of this architecture is its fault tolerance. If an agent crashes, the label remains on the issue. When the agent recovers, it picks up exactly where it left off. No volatile state is lost. Moreover, humans can intervene at any moment without breaking the flow: simply change or remove a label. This turns the system into a true human-in-the-loop, where people and agents collaborate on the same Kanban board.
At Q2BSTUDIO we apply these principles on several fronts. On one hand, we use AI agents to automate repetitive tasks in our process automation projects. On the other, when we deploy solutions on AWS or Azure cloud, agents can manage infrastructure as code, monitor statuses, and react to events without manual intervention. Cybersecurity also benefits: a review agent can detect common vulnerabilities, while a compliance agent verifies that changes adhere to security policies before moving to production.
The combination with artificial intelligence and business intelligence is especially powerful. Imagine a pipeline where a BI agent generates Power BI reports from new data, another agent validates data quality, and a third deploys the updated dashboard. All orchestrated solely with labels. No orchestrators, no extra infrastructure, no blind spots.
This approach also facilitates experimentation. You can test a new agent simply by adding a label and a prompt. If it doesn't work well, remove the label and the flow reverts to the previous state. No need to rewrite complex pipelines. It's an agile way to integrate AI capabilities into real production environments.
For teams already using agile methodologies and issue trackers, the learning curve is virtually zero. Developers immediately grasp the concept of 'label = state.' Product owners can see progress in real time without extra tools. Everything is in one place: tasks, conversations, design decisions, and test results.
Of course, it is not a universal solution. It works best when agents have clear responsibilities and transitions are linear or have few branches. For highly complex flows with many forks, some conditional mechanism may be needed. But in most cases, the simplicity it brings far outweighs the limitations.
At Q2BSTUDIO we have seen how this architecture reduces development time, decreases coordination errors, and allows scaling AI agents without increasing operational complexity. We believe it is a natural step toward more autonomous software systems, where artificial intelligence not only helps write code but also actively participates in managing the development lifecycle.
If your team is exploring how to integrate AI agents into your workflows, we invite you to consider this lightweight approach. You don't need a new orchestrator. You only need the labels you already have and the willingness to let agents talk through them. The result is a self-managing pipeline with full visibility and the ability for humans to step in whenever needed.
