About Atlas AI agents

Atlas AI agents respond to your prompts by retrieving, analyzing, and summarizing data from CDF. They can access assets, equipment records, sensor readings, work orders, and documents. They can also search across all of those sources at once and combine the relevant information into a single response.

​

Identify the right use case

Not every workflow benefits from an agent. Good use cases involve questions that require judgment, draw on multiple data sources, and rely on accurate, well-documented data. These dimensions help you assess whether a problem is suitable for an agent before you start building.

​

Accuracy

Language models can give different answers to the same question. Any use case that requires every answer to be exactly right, such as payroll calculations or safety-critical permit approvals, is not suitable for a fully autonomous agent. This does not prevent using agents as part of those workflows, but critical decision points require your review. Workflows where some variation is acceptable, or where you review the output before acting, are more suitable.

​

Risk

Consider what happens when the agent is wrong. Some tasks have a low tolerance for error but also low consequences if something goes wrong, making them suitable for early experimentation. For tasks where an error could affect safety, regulatory compliance, or production, you should review the agent’s output before taking action.

​

Volume and availability

Agents are most useful for high-volume or always-on tasks. Tasks like reviewing maintenance procedures, interpreting inspection reports, or extracting specific parameters from large volumes of vendor documentation are suitable for an agent. Tasks that require 24/7 availability, such as monitoring operational logs or supporting control room questions, also work well.

​

Predictability

Predictable, repeatable processes typically work better with traditional automation. Agents are most useful in open-ended workflows where inputs vary and each question requires different steps to answer.

​

Documentation and data readiness

Agents have no built-in knowledge of your organization. You need to give the agent context through clear documentation that describes the task, the relevant data, and the rules that apply. Use cases that rely on undocumented knowledge are unlikely to work. The data the agent queries must also be accurate and well-structured. Data quality is not optional. Agents can handle specific parts of many real-world processes, even when you still need to monitor the overall workflow. Identifying those parts can be more practical than trying to automate the entire process.

​

Define the agent’s scope and instructions

Each agent should focus on one workflow or question type, such as equipment fault diagnosis, work order triage, or document Q&A. Keeping the agent’s mission narrow makes it easier to test, maintain, and trust. The skills that support that mission can be as detailed as the domain requires. The agent’s instructions define how it behaves in every interaction. Good instructions cover four things. Scope and identity defines what the agent does and what it should decline to handle. Universal rules apply to every question regardless of topic, such as stating assumptions when data is ambiguous. Output expectations describe what a complete response looks like. Boundaries establish what the agent must not do without user confirmation. Workflow-specific knowledge belongs in skills, not instructions. The Agent builder walks you through writing and testing instructions before you publish.

​

Choose a language model

Each agent uses one language model for all its interactions. The model you choose affects how well the agent handles multi-step questions, how quickly it responds, and how much it costs to run. Start with a higher-capability model to establish a quality baseline, then test whether a smaller model meets your requirements once the agent is working well.

​

Equip the agent with the right tools

Each tool you enable adds to the decisions the agent must make. Enable only the tools the agent’s workflow requires. The Query tool is the foundation for most agents. Other tools extend it for time series data, documents, and custom computation. Tools that work with time series or documents rely on the Query tool to find the right asset or file first, so configure these combinations as pairs and document the expected order in your skills.

​

Build focused skills

Skills give the agent domain knowledge for specific workflows. Each skill contains vocabulary mappings, data model guidance, and tool sequencing instructions, and loads only when the user’s question is relevant. This keeps the agent’s context focused on what matters for the current question. For example, a root cause analysis agent might use three separate skills: one for gathering relevant sensor data, alarms, and work order history; one for identifying patterns and anomalies in that data; and one for building the causal chain from symptom to root cause. Each skill can be tested and improved on its own, and all three could be reused in an equipment health agent without modification.

​

Delegate with subagents

When one agent cannot cover every workflow without losing focus, use subagents (Preview). An orchestrator agent routes a task to another independently configured agent that runs with its own language model, instructions, skills, and tools. Requires runtime version 1.3.0 or later.

​

Manage context and sandbox

Longer, more complex tasks, such as analyzing data across thousands of rows or time series data points, can run into a language model’s context limit or need a safe place to execute commands. Runtime version 2.0.0 (Private preview) adds to-do list tracking, a sandbox execution environment, automatic context compression, and tool result offloading, so agents can handle these tasks without extra configuration.

​

Test early and iterate

Define test cases before you start tuning the agent. A set of 10 to 20 representative questions with specific expected responses gives you a baseline to measure against. Run your test cases after every meaningful change to skills, instructions, tools, or the language model to catch problems early.

​

Trace and debug agent behavior

Testing shows you what went wrong. Tracing shows you why. Every agent response includes a reasoning field that shows which tools the agent called, in what order, and how it used the results. Inspecting this chain is important for agents that call multiple tools in sequence, where a problem in one step can produce a final response that looks correct but is based on incomplete or wrong data.

​

Query the agent

Once published, you can query the agent through the Atlas AI interface or call it through the API and SDK. Specific, context-rich questions produce better results than vague ones.