Skip to main content
Atlas AI agents use tools to retrieve data, perform calculations, and interact with external systems. A language model can reason about a question and decide what to do, but it needs tools to act on those decisions. When the agent calls a tool, it passes structured inputs, receives a result, and uses that result to build its response.

​
Why the right set of tools matters

Each tool you enable adds to the decisions the agent must make. More tools means more opportunities for the agent to choose the wrong one. An agent with a focused tool set makes more reliable decisions than one with a long list of overlapping options. Enable only the tools your agent’s workflow requires. Use skills to document which tool to use and when, so the agent does not have to infer the correct choice at runtime. Some tools work as pairs. Time series tools and document tools need the Query tool to run first. The agent uses the Query tool to identify the right asset or file in the knowledge graph, then uses the specialized tool to retrieve or analyze its data. When you enable these tools together, add sequencing instructions to your skills so the agent knows which tool to call first. For example, if a user asks whether a bearing temperature is trending upward, the agent first uses the Query tool to find the relevant time series in the knowledge graph. It then retrieves the readings with Query time series data points and calculates the trend with Analyze time series. Without sequencing instructions in the skill, the agent may skip the first step or retrieve readings for the wrong time series.

​
Use tools for calculations

Language models are strong at reasoning and language understanding. They are not reliable for exact arithmetic, date calculations, or precise comparisons. When a question has a correct, verifiable answer, a tool should compute it, not the language model. This table shows common question types and the tool that produces a reliable result.
Question typeRoute to
”How many work orders are open?”aggregate_instances in the Query tool, which computes the count server-side rather than listing all records and counting them in context
”Is the vibration on Compressor C-205 trending upward?”Analyze time series, which performs statistical analysis rather than asking the language model to estimate a trend
”What percentage of inspections are overdue at each site?”aggregate_instances in the Query tool, or Run Python code for multi-step calculations the Query tool does not cover
”Of those work orders, which ones took longer than planned?”Examine data semantically, which analyzes results already retrieved in the same conversation without re-querying the knowledge graph
”Run a domain-specific calculation that requires custom packages”Call Function, which runs on Cognite’s cloud infrastructure with full package support and longer execution times
You can reinforce this routing behavior in skills. When the user asks “how many,” “what percentage,” or “compare X to Y,” a skill can instruct the agent to use the appropriate tool rather than reason about the answer directly. See Configuring skills for agents for examples.

​
Tool categories

Atlas AI tools fall into three categories. Query tools retrieve data from the CDF knowledge graph. The Query tool is the foundation for most agents: it works across your entire data model without requiring separate configuration per view. The legacy query knowledge graph tools each target a single view. Analysis tools work with data that has already been retrieved or identified. Time series tools retrieve sensor data and perform statistical analysis. Document tools answer questions from files and generate summaries. The Examine data semantically tool analyzes result sets from earlier queries in the same conversation. Integration tools let the agent run custom code or call external services. Run Python code runs a builder-defined script in a WebAssembly sandbox. Call Function calls a deployed Cognite Function on Cognite’s cloud infrastructure. For the full list of tools and their configuration options, see the Agent tools library.

​
Choosing between Run Python code and Call Function

Both tools run custom Python code, but each fits different situations. Run Python code runs a script you define in a sandboxed environment with access to CDF APIs and standard libraries. Use it for lightweight calculations that stay within CDF and do not need external network access or custom packages. Call Function calls a deployed Cognite Function on Cognite’s cloud infrastructure. Use it when you need custom packages, more memory, longer execution times, or access to systems outside CDF. If a calculation fits within the sandbox, Run Python code can be simpler to set up. If it needs anything beyond those limits, use Call Function.
Last modified on June 18, 2026