What Are AI Agent Skills? Capabilities Explained

Tool use in an AI agent is the ability to invoke external functions, APIs, or services as part of completing a task. When an AI agent is reasoning through a problem and determines it needs information or needs to take an action that requires interacting with an external system, it calls a tool - which might be a REST API call to your CRM, a database query, a web search, a file read, or a form submission. Tool use is what distinguishes an AI agent from a conversational AI model: a model produces text in response to a prompt; an agent uses tools to take real actions in the world. Every enterprise AI agent depends on tool use to deliver outcomes beyond generating text.

Yes - an AI agent can write to your database, provided it has been given the appropriate tool wrappers and access credentials to do so. Database write access is one of the most common production agent capabilities: a customer service agent that resolves a case needs to update the ticket status; a loan origination agent that approves an application needs to write the record. Write access is managed through standard access controls - the agent is given the minimum permissions it needs for its specific tasks, and every write action is logged with timestamp, data modified, and the reasoning that triggered the action.

An AI agent's training refers to the knowledge embedded in the underlying language model - the facts, reasoning patterns, and domain knowledge that were learned during the model's training process. This is static: it was determined before deployment. An AI agent's skills, by contrast, are the capabilities added through tool integrations, memory systems, and workflow configuration after training. Skills are dynamic - they can be added, removed, or updated without retraining the model. Training determines what the agent knows; skills determine what it can do.

Skills are determined by mapping each step in the workflow the agent will automate to the data it needs to access, the systems it needs to interact with, and the decisions it needs to make. Start by documenting the complete workflow as a human currently executes it: what information does a person look up, what systems do they open, what do they write, what decisions do they make, and what do they do when they are unsure? Each of those steps maps to a required skill. The full skill map for a production enterprise agent typically includes 6–12 distinct skill categories. Upcore conducts a structured workflow analysis as the first step in every deployment engagement.

Yes - new skills can be added to a deployed AI agent without retraining the underlying model. Adding a skill typically means adding a new tool integration: writing an API connector, a database query function, or a form-fill script, and registering it with the agent's tool execution layer. The agent can then call the new tool as part of its reasoning process. This is one of the key advantages of the tool-use architecture: the agent's capability set is extensible without the time and cost of model retraining. Enterprise agents are regularly updated to add new system integrations and data sources as the scope of the deployment expands.

Human-in-the-loop (HITL) is the agent's ability to recognise when a task is outside its authority or confidence threshold and pause to route it to a human reviewer - along with the context needed for the human to make a decision quickly. It is a skill because it requires active judgment: the agent must evaluate its confidence in a proposed action, compare it against a policy for when autonomous action is appropriate, and trigger a structured handoff if the threshold is not met. A well-designed HITL skill includes detection logic, routing to the right person or queue, a context package summarising what the agent has done, and a mechanism to resume the workflow once the human has responded. HITL is not a fallback for poor agent performance - it is a deliberate architectural feature that defines the agent's operating boundaries.

Yes - code execution is one of the twelve standard skill categories for enterprise AI agents. A code execution skill gives the agent the ability to write and run code within a sandboxed execution environment, typically Python. This is essential for data analysis agents, financial modelling agents, and any use case where the answer requires computation rather than information retrieval. The code runs in a controlled sandbox with appropriate resource limits, and the output is returned to the agent for use in its response. Code execution is added when the use case requires computation, and the sandbox configuration is hardened to match the security requirements of the deployment environment.

Each skill is tested through a structured process covering three areas. First, functional correctness: does the skill produce the expected output given a range of inputs, including edge cases and malformed inputs? Second, integration reliability: does the tool call succeed consistently under load, with appropriate retry and error handling for API failures or timeouts? Third, trust calibration: what is the agent's confidence when it uses this skill, and at what confidence threshold should it escalate to a human rather than act autonomously? Trust calibration is particularly important for skills that write to production systems or make decisions with downstream consequences. Upcore runs a dedicated pre-production testing phase for every skill in the agent's toolkit.