How does an AI agent work?
An agent runs a loop. The application gives the model a goal, instructions and a list of tools; the model picks a step; the application runs it; and the result goes back to the model for the next decision. NIST describes an LLM-based agent in these terms: a software system that repeatedly prompts a model, processes its outputs, for example by calling a function the model selected, and feeds the results back as part of the next prompt. Agents may also have memory or planning.
Anthropic draws a useful line between workflows, where code fixes the order of model calls and tools, and agents, where the model directs its own process and tool use. In its words, agents are “typically just LLMs using tools based on environmental feedback in a loop”.
Take the question “Why was my invoice higher this month?” An agent with the right tools might fetch the invoice, list recent plan changes and explain the difference. An agent on a help page might need only one tool: a search of the product docs.
What makes a voice agent an AI agent?
A website voice agent is an AI agent whose main job is the conversation. Tools let it look things up or note what the caller said. Each extra tool widens what it can do and what can go wrong: an agent that can issue refunds needs identity checks, permissions and confirmation steps, while one that can only read approved docs mainly needs good docs and testing.
The word “agent” also describes software that operates other software. A coding agent connected through the Model Context Protocol (MCP) can change a product’s settings by calling its tools, which is a different role from the agent your customers talk to.
What can Kawl Now’s agents do?
Kawl Now gives each workspace two AI agents, one for support and one for sales. Each has its own approved knowledge, instructions, greeting, voice, widget and allowed website origins.
During a call, an agent can do two things besides talk: search the approved knowledge for its channel, and record a bug report, feature request or other feedback when the caller asks. It can’t look up or change customer accounts, book meetings or update a CRM, so it suits product questions rather than account changes.
The agent has no tool to transfer a call. Human handover, an early release, works through the widget instead: once your team verifies a UK or US phone number and turns handover on for an agent, a caller can choose Talk to a human, and whoever answers hears a private briefing before the caller is connected. Customer routing is still being tested, so keep your usual contact options.
You can also manage Kawl Now from your own coding agent. A remote MCP server and a REST API take a workspace token, and the MCP tools cover settings, knowledge, approval, conversations, automatic insights and the embed code. A token can approve knowledge as well as add it, so decide whether your coding agent should.
Common questions
What is the difference between an AI agent and an AI workflow?
In a workflow, your code decides the steps and calls the model at fixed points. In an agent, the model decides which step or tool comes next, based on what it has found so far. Workflows are more predictable, and agents suit open-ended tasks that can’t be scripted in advance.
Can a Kawl Now agent take actions in my product?
No. A Kawl Now agent answers from approved knowledge and can record feedback on the conversation. It has no access to your product or your customers’ accounts, and there are no custom tools or webhooks to give it any.