With the help of AI Agents and Model Context Protocol (MCP) for tool access, organizations building enterprise AI solutions can get several steps closer to a fully autonomous system. Recent developments in open-source autonomous agents (like OpenClaw, Ralph Wiggam loops) have proven that AI is no longer just a chatbot waiting for a prompt; it is an active participant capable of executing complex workflows. While this area is evolving very fast, the technology is mature enough to be termed an early-age era for agents, much like TCP/IP was for the web.

From Alerts to Autonomous Investigations

Traditional SOAR platforms are very good at orchestrating playbooks and routing alerts. However, they require human analysts to perform the task of context gathering and decision-making. AI Agents are a perfect foil here since they can execute complex natural language investigation workflows and can even help run critical containment steps without waiting on slow manual intervention, forming the backbone of modern AI enabled enterprise security solutions.

Multi-agentic systems work on one core concept. Creating a shared state contract, defining how each property is going to evolve over time (as execution proceeds), and which agents can modify which properties of the shared state. Typically, an agent is only responsible for delivering a “delta” in the state contract, upon which some actions/analysis can be performed.

Taking architectural inspiration from recent agentic breakthroughs, a robust AI Agent is defined by four key characteristics:

• Planning: The agent divides the larger complex execution plan into smaller steps, enabling it to execute the plan reliably and stay on track. Several research based solutions use this technique to build multi-agent systems. You would often see some sort of “planner” agent in the mix.
  
  
• Memory & State Persistence: Helps AI Agents make informed and precise decisions based on gathered context. Modern agents maintain “persistent memory” (often as plain-text diaries or structured state files) so they can remember user preferences and past alerts across long-running sessions. Fun fact: Openclaw stores its guidelines and skills as simple markdown or txt files in your local persistent storage. We can enhance and store these in a shared state (JSON/YAML) contract for our agents as part of a scalable end-to-end AI solution.
  
  
• Proactivity (The “Heartbeat”): Unlike a standard LLM that sits idle until you type a message, autonomous agents operate on a continuous loop or “heartbeat.” They wake up periodically to scan SIEM logs, check task queues, and trigger investigations proactively.
  
  
• Tool Use (Skills): Interfaces with external APIs or services to gather information or retrieve logs. The arrival of MCP has made interfacing with these external data sources modular, reliable, and much easier to implement.

The Core Engine: Model Context Protocol (MCP)

At the core of each AI Agent is MCP. MCP is a standardized protocol for providing context to LLMs via on-demand tool execution, acting much like the “Skills” or plugins that allow modern agents to connect to local environments. Here is what MCP provides us, which can be used to extract:

• Context Retrieval: AI Agents can query relevant data sources (SIEM logs, Threat intelligence platforms, EDR platforms) using natural language instructions, so they can be used by both the AI specialists and the autonomous agents without the need to change SOAR playbooks each time a new data source is added or removed.
  
  
• Context Enrichment: AI Agents can use the context (data) provided by each platform MCP Server to run their investigation tasks and add more context to the base alert.
  
  
• Action Execution: Based on the context of the alerts and the fetched context that each agent has, agents can automatically perform steps to stop the actual attack from spreading.

When Do We Need Multi-Agent Systems?

While a single AI agent can perform a specific task like detecting suspicious logins, it operates within a limited scope. Multi-agent systems take this further by enabling multiple specialized agents to collaborate, share context, and coordinate actions across domains, creating a far more robust and scalable system often seen in advanced enterprise AI solutions.

Below is one simple design that we created to demonstrate the use of AI in SOC operations and as part of a broader AI enabled enterprise security solutions architecture.

As part of this system, there are several specialized agents operating in tandem. Below is the expertise of each:

• Orchestrator: Holds the information of all the sub-agents at its disposal and their duties. Responsible for delegating tasks to appropriate agents in order to keep the system loosely coupled.
  
  
• SOC L1: Performs the preliminary investigation tasks, such as searching for data through SIEM tools or retrieving the threat information about IOCs from Threat Enrichment platforms.
  
  
• Notifier: Notifies/escalates the respective incident based on the preliminary investigation reports from the SOC L1 agent.
  
  
• Report Generator: Generates a summary of the investigation the AI Agent has done, detailing the tools called and decisions taken at each step.