Agents

Define the role

Describe the agent’s purpose, relevant knowledge, and limits. A specialized label does not guarantee expertise, factual accuracy, or permission to perform actions.

Choose capabilities deliberately

A useful agent needs appropriate model and tool access. Skills describe procedures; tools expose governed operations.

Optional presentation and connections

The agent interface includes voice-assignment and correspondence-related configuration. Availability depends on the enabled services and operator setup; these are not blanket offline capabilities.

Validate the intended work

Use representative tasks to check behavior and boundaries before relying on an agent. Specific preconfigured expert offerings should be listed only with release-specific evidence.

Choose an Agent and inspect the reply

Select an Agent suited to your task where the selector is available. Review the active Agent and model attribution shown with the conversation or reply. A requested choice can be affected by routing and capability rules. Test the role with a small, verifiable task before relying on it.

Ask Friday to guide you

In an authorized FridayLocalAI conversation, try:

Help me choose among the Agents I can see: [names and descriptions]. My task is [goal]. Explain the tradeoffs, identify any required capabilities, and suggest a small test. Do not assume any Agent has access to my files.

Tell Friday what you can see and which step you have reached. If the instructions do not match your screen, say so; ask for the relevant Help topic or contact your operator. Friday cannot be assumed to see your screen or know your current settings.

Understand the role before selecting it

An Agent is a configured assistant role within the application. A useful role has a clear purpose, such as helping with a kind of analysis or a recurring workflow. Its description can help you decide whether it suits the task, but a name alone does not guarantee subject expertise, factual accuracy, or access to the resources needed to complete the work. Read the available description and compare it with your intended outcome. If the fit is uncertain, ask Friday to explain the choices you can see. Supply the relevant labels and your goal rather than expecting the assistant to know every available configuration automatically. Begin with a small task whose result you can evaluate before making the Agent part of an important or repeated workflow.

Keep Agent identity separate from the model

The Agent’s role and the model producing a response are different parts of the system. An Agent may guide how work is approached, while a model supplies the response capability. FridayLocalAI exposes requested and active attribution so users can inspect the response path rather than assuming every choice maps directly to a single model in every situation. When comparing behavior, record the active Agent and model information shown with the reply. Keep the prompt and source material similar enough that the comparison means something. A different result may come from changed context, routing, or input as well as from the role itself. Avoid assigning all differences to the Agent name without checking what actually happened during the request.

Provide a task brief

Even a specialized Agent needs a clear request. State the goal, relevant context, constraints, and expected output. For a review task, explain what should be checked and what evidence is available. For a drafting task, identify the audience and the form of the result. If important information is missing, ask the Agent to question you before proceeding. For example, a request to compare two approaches should specify the criteria that matter. Otherwise, the answer may choose criteria that are plausible but irrelevant to your decision. A useful brief also identifies what should not happen, such as sending a draft or modifying a source file. This makes the intended boundary visible without pretending that the prompt itself grants or removes runtime permissions.

Verify source and Tool access independently

An Agent can discuss a file without being able to read it, and it can describe an operation without being able to perform it. Confirm source availability with a checkable question and confirm Tool availability through the appropriate interface or operator guidance. A role description should not be treated as evidence that every necessary capability is enabled. If the task requires a connected provider or external action, understand that boundary before supplying sensitive material or authorizing the action. Optional voice or correspondence-related configuration does not establish blanket offline functionality. Ask what the current deployment supports and what remains unverified. The useful result is a clear match between the role, the task, the available resources, and the permissions governing their use.

Evaluate the role with representative examples

Choose a few small examples drawn from the kind of work you expect to repeat. Include a straightforward task, a task with missing information, and a case that tests an important constraint. Define how you will assess usefulness before reading the responses. Look for factual accuracy, relevance, adherence to constraints, and appropriate handling of uncertainty. Keep the evaluation proportionate. You do not need a large benchmark to discover that an Agent consistently misunderstands the desired format. On the other hand, one successful example does not prove broad reliability. Record the result and adjust the task brief or consult the operator as appropriate. If configuration changes, repeat the relevant example rather than assuming the earlier result still applies.

Maintain the human review step

Specialization is helpful when it reduces the effort needed to begin a task or follow a useful process. It does not remove the need to inspect important facts, sources, and artifacts. Ask the Agent to explain its assumptions and identify what should be verified. For completed operations, examine the actual result instead of accepting the role’s confidence as evidence. If the available Agent does not fit, you can still ask for a bounded explanation or a draft within its capabilities. Requests to create or modify Agent configuration should follow the controls and permissions established for the deployment. Early-access participants should work with their operator on those choices. The objective is a role you can understand and evaluate, not a label that encourages you to assume the application can do more than has been configured and verified.