Model completions

Calling a Llama model for a written answer is the core move here. When a task lands, say a support reply or a summary, the agent assembles the prompt from your data, sends it to the Llama or open-source model the account has access to, and reads back the generated text. It then applies an approved rule to place that text where it belongs, a draft field, a ticket, or a document. Answers that come back empty, run past a length limit, or touch a topic the rule marks as sensitive are held for a person to approve first.

Response filtering

Not every model answer should ship untouched. After a completion returns, the agent reads the text against a customer-approved checklist, looking for signs the rule flags: refusals, off-topic drift, or content that names something it should not. Clean responses are written straight into the target record so work keeps moving. Anything that trips the checklist is set aside with the reason attached and routed to a reviewer, who edits or rejects it. The agent never overrides that judgment; it only prepares the material and records which answers passed and which needed a human.

Prompt template selection

Different jobs need different prompts, and an agent can pick the right one. Reading the incoming task type and any labels on the record, it selects a stored prompt pattern, fills in the specific values, and calls the model through the account's Llama endpoint. The chosen template and inputs are saved alongside the result so the run can be traced later. When a task type has no matching template, or the values needed to fill one are missing, the agent stops and asks a person which pattern to use rather than guessing at the setup.