Building the coordinated AI workforce
The first generation of enterprise AI has largely focused on the capabilities of individual models and agents.
Can an agent investigate an alert? Analyze a vulnerability? Gather threat intelligence? Write a detection? Prepare a report?
These are useful questions, and specialized agents can create meaningful value. But they do not address the larger operational challenge. Real work rarely fits within the boundaries of one agent.
A security investigation may require information from identity, endpoint, network, cloud, vulnerability, and threat-intelligence systems. It may involve gathering evidence, correlating activity, assessing risk, documenting conclusions, recommending action, and escalating a decision. No single agent will necessarily have the context, tools, permissions, or expertise required to perform every part of that work effectively.
The future of enterprise AI will therefore require more than increasingly capable individual agents. It will require a system through which specialized agents can organize around an objective, divide responsibility, work concurrently, and coordinate their contributions into a reliable outcome.
It will require an AI workforce.
A collection of agents is not a workforce
Enterprises are beginning to adopt copilots and specialized agents across different functions. Each may address a particular task, application, or data source.
The result can be useful local automation. It can also create a new form of fragmentation.
Each agent may have its own context, its own permissions, its own tools and integrations, its own memory, its own operating procedures, its own logs, and its own approach to human oversight.
Adding more agents does not automatically make the overall operation more effective. Without coordination, it can create more outputs for people to interpret, reconcile, and govern. This is similar to building a human organization by hiring capable individuals without defining how they will communicate, assign work, share information, or remain accountable for a common result.
Individual capability matters. The operating model matters just as much.
Work should be organized around the objective
Many multi-agent systems use a central orchestrator to divide a request into predefined tasks and assign those tasks to agents with known roles.
That approach works when the required process and available capabilities are already understood. It becomes more difficult when the objective is novel, conditions are changing, or the best combination of capabilities cannot be determined in advance.
A more adaptive system allows agents to evaluate the objective and indicate how their capabilities apply to the work being requested. The process can work as follows:
- A person states an objective.
- Available agents evaluate the request.
- Agents identify how their capabilities relate to the current objective.
- The system selects participants and assigns responsibilities based on fit.
- The agents contribute to the plan.
- Multiple agents perform work concurrently.
- Their activities and results are coordinated into an overall outcome.
- While performing their assignments, agents may create local plans and coordinate supporting work by other agents, which may do the same in turn.
The distinction is important. The system does not simply route a predetermined task to a predetermined agent. It forms an appropriate team around the objective and lets that team help determine how the work should be completed.
The organization can also form recursively. An agent contributing to a larger objective may develop its own local plan, identify supporting work, and coordinate other agents capable of performing it. Management and contributor roles are therefore temporary and task-specific: the same agent can contribute upward to a broader objective while coordinating work beneath it.
Consider one example: a threat hunt.
An analyst tells Bricklayer, through conversation, to hunt for and report on QTFY, a state-sponsored threat actor.
Bricklayer then determines the work required and assigns the available Public Knowledge, Qualys, Splunk, and Reporter agents to specific tasks, which they perform concurrently. Insight groups are created automatically along the way, structuring context for the tasks that depend on it. The Reporter agent then assembles the findings into a multi-section report.
Coordination creates leverage
Capability-aware coordination can improve the way an AI workforce operates in several ways.
- Better task assignment. Work can be directed to the agents best suited to perform it in the context of the current objective.
- Parallel execution. Multiple areas of the problem can be addressed simultaneously rather than through a strictly sequential workflow.
- Greater adaptability. The system can respond to objectives that were not completely anticipated when the agents were created.
- Easier expansion. New agents and capabilities can be introduced without redesigning every existing workflow.
- Shared responsibility for an outcome. Agents contribute to the broader objective rather than producing isolated responses that a person must manually assemble.
- Nested autonomy. Each agent can determine how best to satisfy its assignment, including recruiting and coordinating supporting agents, while remaining accountable for the result it returns to the broader team.
This does not eliminate orchestration. It distributes orchestration across the workforce. Instead of relying exclusively on one permanent coordinator, agents can assume temporary management responsibility for the portions of work they are best positioned to organize.
Coordination is only one part of the AI harness
A coordinated group of agents still cannot be placed into enterprise operations without a broader set of controls.
Organizations must determine what information each agent can access, which tools and credentials it can use, what actions it is authorized to perform, which procedures it must follow, when human approval is required, how its behavior will be tested, how decisions and actions will be observed, what happens when an agent encounters an exception, how performance and business outcomes will be measured, and how agent changes will be versioned, promoted, and rolled back.
These capabilities form an AI harness around the agents. The models provide intelligence. The agents provide specialization. The harness provides the context, coordination, and control required to apply them to real work.
At Bricklayer, we think about that harness through three architectural pillars:
Context. Agents understand the organization, environment, and work.
Coordination. Agents organize around a shared objective, contribute according to their capabilities, and coordinate supporting work through temporary, task-specific relationships.
Control. People remain accountable for autonomous execution.
All three are required. Context without coordination leaves capable agents working in isolation. Coordination without control creates operational and governance risk. Control without sufficient context limits the quality of the work.
Why cybersecurity is an important proving ground
The operating model behind Bricklayer is not specific to security. The same pattern of agents evaluating an objective, reporting their fit, and working concurrently under a shared plan applies wherever work is distributed across specialized capabilities.
Cybersecurity is where we chose to develop and validate it first, because it provides one of the most demanding environments available.
Security work is distributed across tools, data sources, teams, and areas of expertise. The environment changes continuously. Decisions can be consequential. Evidence must be preserved. Actions must be governed. Exceptions are common, and human accountability cannot disappear.
These characteristics make cybersecurity difficult for isolated agents. They also make it a useful proving ground for coordinated AI workforces: a model that holds up here, under novel objectives, real stakes, and strict governance, is one that can extend into other domains with confidence.
The objective is not simply to generate a faster answer. It is to complete a meaningful portion of operational work safely, reliably, and with a clear record of what happened and why. That requires agents to work across the operation while remaining within defined procedures, permissions, and escalation boundaries.
Protecting an early part of the vision
Bricklayer began developing this approach before multi-agent coordination became a common part of the AI platform discussion.
We are pleased that the United States Patent and Trademark Office has now provided a Notice of Allowance for a Bricklayer AI invention titled “Systems and Methods of Artificial Intelligence Agent Coordination and Management.” It joins other approved Bricklayer patents covering agentic policy enforcement, the mechanisms through which agent behavior is governed and controlled in production.
The USPTO examiner specifically recognized the distinction between conventional orchestration and a system in which agents report how their capabilities apply to the current query. That information allows work to be assigned based on fit and enables agents to participate in planning while contributing to the work itself.
The broader architectural idea is that these roles need not be permanent. An agent can contribute to an assignment while also coordinating supporting work by other agents, allowing the workforce to organize recursively as the objective unfolds.
The Notice of Allowance is an important milestone, but the larger story is not the patent itself. The larger story is the change taking place in how work will be organized.
From software tools to AI teams
Enterprise software has traditionally provided tools for people to use. The next phase will include agents that can perform portions of the work themselves.
That transition will not be achieved by placing a separate agent inside every product and asking people to coordinate the resulting outputs. Enterprises will need a common operating layer through which agents can understand the environment, identify where they can contribute, work with other agents and people, follow organizational controls, and remain accountable for measurable outcomes.
The companies that succeed with agentic AI will not simply have access to the best individual models or the largest number of agents. They will know how to turn those agents into a coordinated workforce that can form around an objective, organize recursively, and adapt its structure as the work unfolds.
