Building Halo's multi-agent architecture: from one loop to five specialists

The evolution of artificial intelligence-based systems towards multi-agent architectures is transforming how companies approach complex problems, from cybersecurity to process automation. By dividing a monolithic workflow into specialists with clearly defined roles, greater robustness, scalability, and ease of maintenance are achieved. This approach, which we have seen applied in custom application development projects, allows each module to be refined independently, accelerating development cycles and reducing technical debt.

In practice, the separation of responsibilities is key: some agents handle exploration and reconnaissance (such as port scanning or service enumeration), while others execute offensive or exploitation actions. Orchestration and planning agents, on the other hand, remain as purely logical layers, without direct access to tools. This not only avoids unwanted side effects but also facilitates the incorporation of new modules without rewriting the system's core. Such a design is especially relevant in environments where AI for businesses must integrate with cloud and on-premise platforms, maintaining standardized communication between all components through a common message format.

Building this type of architecture usually begins with stubs or simulations of the specialized agents, allowing the orchestration flow to be validated before connecting real systems. Once the real modules are ready, the exchange is transparent: a change in the routing configuration is sufficient. During tests against real targets, challenges arise such as parameter errors or interpretation failures, which a well-designed debugger agent can diagnose with high confidence, recommending corrective actions without human intervention. This demonstrates how artificial intelligence and AI agents can collaborate to close the feedback loop autonomously.

From a business perspective, this architecture fits perfectly with modern cybersecurity needs, where penetration testing and vulnerability detection require speed and precision. Additionally, the use of AWS and Azure cloud services provides the necessary scalability to run multiple agent instances in parallel, while business intelligence services tools like Power BI can visualize the aggregated results of testing campaigns. At Q2BSTUDIO, we help organizations design and implement these solutions as part of a custom software ecosystem, ensuring that each piece aligns with business objectives and security best practices.