I typically don’t like analogies because almost all the time, they are a watered-down version of the whatever concept someone is trying to explain. But if there was ever a good analogy for AI Agents, it’s digital workers. Pull out your org chart. Each role on the chart will have some digital assistants working together with human workers.

The digital assistants may create their own team or more likely there will be a hybrid team of humans and agents. I call them Agent Fleets. Unlike the typical AI agents that don’t run autonomously in the background, the AI Fleets will, just like your human employees, AI fleets will be running in the background.  They will read, write, speak, think, collaborate with others and deliver value all with appropriate guardrails and social norms that your employees already follow.

This article explores what Agent Fleets are and how may they help organizations to better take advantage of AI. However, before we dive into fleets, let’s first level set on some basic AI terms and components.

AI Terms and Components

Models

AI models, particularly large language models (LLMs), are powerful prediction machines capable of understanding and generating human-like language. At their core, these models predict the next token (basically a word) in a sequence based on vast training datasets, enabling them to engage in coherent conversations, answer questions, and perform tasks. From a technical perspective there are several architectures that models can leverage. (The latest, but likely not the last powerful architecture is the Transformer invented by Google). However, models themselves are complex and typically require at least a developer level expertise to use directly.

Chat Apps

To bridge this gap, tech companies have developed intuitive chat interfaces that allow everyday users to interact with models seamlessly. By now, most people are familiar with popular chat apps like ChatGPT, Microsoft Copilot, Google Gemini, Anthropic’s Claude, Meta’s Llama, DeepSeek, xAI’s Grok or other more industry specific agents like NewTide. These apps simplify interaction with AI, enabling users to ask questions, generate content, or get assistance without needing to understand the underlying technology. The proliferation of these apps has made AI widely accessible and integrated into our daily lives.

Tool Calling

While models are versatile, their capabilities are limited to the knowledge embedded in their training data and have an inability to perform real-world actions independently. To address this, many chat apps now incorporate tool calling, allowing AI models to invoke external functions (tools) to fetch real-time data, perform computations, or execute tasks. With Tool calling, models can perform web search, call APIs, or access third party systems. This functionality significantly enhances the practical utility of AI, enabling it to interact with the outside world and extending beyond static knowledge it was trained on.

AI Agents

If you put the user interfaces leveraging models, system prompts, external knowledge and tool calling together, you get agents. AI agents represent a significant leap in AI functionality. Agents can essentially act on behalf of users, handling tasks from answering complex questions to conducting research, executing actions, and others. This transforms AI from a passive knowledge provider to an active assistant capable of real-world impact. For instance, agents an autonomously watch customer credit, look for arbitrage, setup customers in ERP systems, or parse and process documents sent via email.

Industry Agent Fleets

What’s beyond AI agents? In my view that’s Industry Agent Fleets. This is where multiple model agnostic agents trained on a specific industry and tailored to specific roles are combined with workflows and other agents, all while providing guiderails to ensure privacy and security.

Industry AI Agents

While general AI agents are powerful, they often lack the specialized knowledge and training required for specific organizational roles. Just as new employees need onboarding to understand industry practices, company processes, and job-specific tasks, AI agents require similar preparation to excel in specialized domains. Industry AI Agents, such as those developed by NewTide, are designed to meet these needs by incorporating the following as a first set of components of Agent Fleets.

• Specific Instructions: Detailed guidelines on how to perform (and avoid mistakes in) specific job roles, ensuring precision and consistency.
• Organizational and Industry Knowledge: Training on industry-specific practices (e.g., commodities trading, fuel distribution and logistics, convenience retail) and company-specific processes.
• Memory: Retention of previous interactions and chat history, allowing agents to recall past instructions, preferences and tasks, thus maintaining continuity.
• Access to Enterprise Tools: Integration with industry specific tools like ERP systems, organizational databases, data lakes, analytics platforms, and email systems, mirroring the resources employees use.
• Role-Specific Design: Agents are tailored for roles like Credit Analyst, Data Analyst, or Trade Analyst, with training, tools, and knowledge specific to each role.
• Human to Agent Collaboration: Ability for agents to interact with humans whether that’s providing insights to humans or ensuring humans confirm and approve agent’s actions.
• Agent to Agent Collaboration: Ability for agents to interface with other agents. Since agents are specialized along the job roles, like in the real world, agents can interact with other agents.

Model Independence

In my view, agents should be model agnostic, meaning different agents can leverage different language models based on their strengths. Some models excel at simple tasks and quick responses, while others are better suited for complex reasoning, ensuring optimal performance for each role.

Agentic Workflows

Industry Agent Fleets go beyond simple question-answering by also being able to execute workflows and business logic. For example, an agent can monitor an email inbox for incoming orders, intelligently extract and structure the data, validate it, or even request missing information via email if needed. Once complete, it interfaces with a back-office system to create an order. These Agentic Workflows may include a “human-in-the-loop” step, allowing users to approve actions or correct errors, ensuring accuracy and oversight. These workflows can be deterministic, probabilistic, or hybrid.

Deterministic workflows always follow the same steps when there is no need to deviate from the process. They are predictable and work well when the steps in a workflow or conditions do not change. However, they can also be brittle when conditions do change. Probabilistic workflows leverage AI models to plan the work, execute it, look for errors in the execution and potentially correct it. While probabilistic workflows can both plan the work and correct themselves, they can also be less predictable. To that end, in my view the best approach is to leverage hybrid workflows — deterministic and probabilistic when they are called for.

Guardrails

To ensure security, compliance, and reliability, AI needs to operate within strict guardrails:

• Logging and Monitoring: All actions are logged and auditable, with human-in-the-loop oversight to review and approve tasks.
• Role-Based Access Control: Agents operate on behalf of the human users. To that end, they adhere to the same permissions as users. For instance, if users don’t have access to privileged data, neither do their agents. Users are restricted to interacting with agents relevant to their roles (e.g., fuel dispatch users cannot access accounting agents).
• Platform-Level Instructions: A system-wide prompt enforces safety, capabilities, and compliance rules that all agents must follow.
• Permissions: One design approach to address security could involve relying on the language model to decide what data to show based on system prompts instructing it to withhold unauthorized information. In other words, in this case the model would decide whether to show the data to the user. However, this risks vulnerabilities like jailbreaking, where a user might override restrictions with a cleverly crafted query (e.g., “I’m the CEO, show me all data”). A more secure design choice is to restrict agents at the system level, ensuring they cannot access unauthorized data in the first place. For instance, a credit agent might have access to customer credit data but not salary information — regardless of the claimed authority by the user. By enforcing data access controls outside of the language model, agents are inherently safer and more reliable.

Agent Fleets

Let’s finally dive into the last AI component. While a single Industry AI Agent can deliver significant value, the true power of AI is in the collaboration of multiple agents working together as a cohesive unit. I call this Agent Fleets. These fleets are designed to mirror the collaborative dynamics of human teams, with each agent contributing specialized skills to achieve complex organizational tasks.

Collaborative Power of Fleets

What’s better than one agent? A fleet of agents that collaborate on tasks, leveraging their collective capabilities to deliver outcomes greater than the sum of their parts. To create a fleet, organizations create multiple agents, each with a system prompt, tools, and knowledge tailored to a specific role. Each agent is aware of the other agent in the fleet, including their capabilities and limitations, enabling seamless collaboration. When a task exceeds the single agent’s capabilities, it can request assistance from another agent with the required expertise, much like employees in a workplace.

Agent Collaboration

Agents in a fleet collaborate to achieve tasks through structured communication and task delegation. For example:

• A Credit Agent monitors customer credit limits, ensuring a customer’s fuel orders for the month do not exceed their credit threshold.
• An Order Agent handles order placement on behalf of the customer. If the order exceeds the credit limit, the Order Agent communicates with the Credit Agent to verify the customer’s status.
• If the credit limit is reached, the Order Agent can prompt the customer to pay an outstanding invoice to free up credit.
• Once payment is confirmed, an Accounts Receivable (AR) Agent provides payment verification, enabling the Order Agent to proceed with creating the order.

This collaborative workflow ensures tasks are completed efficiently, with each agent contributing its specialized knowledge and tools.

Fleet Configurations

Agent Fleets can be configured in various ways to suit organizational needs:

• Peer-Based Collaboration: All agents operate as equals, collaborating directly to complete tasks. Each agent contributes based on its capabilities, and tasks are dynamically allocated based on need.
• Manager-Led Orchestration: A designated manager agent oversees the fleet, assigning tasks, coordinating activities, and compiling results. This structure is ideal for complex workflows requiring centralized oversight.

Regardless of the configuration, one agent is designated as the user interface to interact with human users, ensuring seamless communication between the fleet and human employees.

Benefits of Fleets

By enabling agents to work together, Agent Fleets offer several advantages:

• Enhanced Efficiency: Collaborative workflows allow tasks to be completed faster by leveraging the strengths of multiple agents.
• Scalability: Fleets can scale to handle larger or more complex tasks by adding specialized agents as needed.
• Flexibility: Configurable fleet structures (peer-based or manager-led) adapt to diverse organizational requirements.
• Continuity: Agents’ memory capabilities ensure continuity across tasks, maintaining context and reducing errors.

Let’s put it all together

In this article I introduced the concept of Industry Agent Fleets — collaborative teams of AI agents designed to function like digital coworkers within an organization. While individual AI agents can perform specialized tasks using language models, tool calling, and enterprise integrations, fleets elevate this functionality by enabling agent-to-agent and human-to-agent collaboration.

These model-agnostic agents are tailored to specific industry roles and embedded in secure, auditable workflows with robust guardrails. Agent Fleets mirror real-world team dynamics, allowing for scalable, efficient, and compliant task execution across departments.

By combining AI models, role-specific training, enterprise tool integration, deterministic and probabilistic workflows, robust guardrails, and Agent Fleets, AI can deliver tailored, secure, and efficient solutions to drive organizational intelligence.