What are AI agents?
AI agents are software systems that pursue a goal by planning steps, calling tools, observing the results, and adjusting until the task is done or a human steps in. The reasoning engine inside most current agents is a large language model (LLM). The model reads the goal, decides the next action, and emits a structured call to a tool such as a database query, an API, or another agent. The platform runs that call, feeds the result back into the model’s context window, and the loop continues.
The loop is what separates an agent from a chatbot. A chatbot answers a message and waits. An agent keeps working across many steps, holds state between them, and acts on other systems. A customer service agent, for example, can look up an order, check the carrier’s tracking feed, and open a replacement order from one request.
How AI agents work
- Goal and context. The agent receives a task, a system prompt that defines its role and limits, and access to relevant data such as the customer record or the open ticket.
- Planning. The model breaks the task into steps. Some agents write the plan out in full; others decide one step at a time, as in the ReAct pattern.
- Tool use. The model issues a tool call with structured arguments. The platform executes it against the real system and returns the result.
- Observation. The agent reads the result, checks it against the goal, and decides whether to continue, retry, or stop.
- Completion or escalation. The agent returns a result, or hands off to a person when it reaches an approval point, an error, or the edge of its permissions.
Agent memory, tool governance, and an approval workflow sit around this loop. Without them, an agent is a model with unrestricted access to whatever the integration allows.
Why AI agents matter
Agents move AI from answering questions to completing work. A model that can only write text still leaves the copying, clicking, and checking to a person. An agent that can read the ticketing system, query the warehouse database, and post a note in the CRM removes those handoffs. The business value shows up as cycle time: a quote that took two days takes an hour, and a reconciliation that waited for month-end runs nightly.
The same capability raises the stakes. An agent with write access to a system of record can create a purchase order, change a price, or close a case. Every deployment therefore needs an identity for the agent, least privilege on each tool, a log of every action, and a human approval step wherever the agent can change a record.
AI agents in practice
A regional distributor runs an agent on its inbound order desk. Customers email purchase orders in many formats. The agent extracts the line items, matches part numbers against the item master, checks stock and pricing in the ERP system, and drafts a sales order. If every line matches and the customer is in good standing, the order posts. If a part number is ambiguous or the credit limit is exceeded, the agent routes the draft to a customer service representative with a note explaining the problem. The representatives still own the exceptions, but they stop retyping clean orders.
In Kirk Tech Solutions engagements, agents of this kind usually run on a private AI platform with an open-weight model, so purchase orders and customer records never leave infrastructure the company controls. FlatClaw, the open-source private AI platform from Kirk Tech Solutions, runs agents this way, with per-user identity and approval steps built in.
AI agents compared with chatbots and workflow automation
| Property | Chatbot | Workflow automation | AI agent |
|---|---|---|---|
| Decides the next step | No, it answers one message at a time | No, the steps are fixed in advance | Yes, from the goal and the results so far |
| Acts on other systems | Rarely | Yes, through fixed connectors | Yes, through tools it chooses at run time |
| Handles unexpected input | Asks the user | Fails or routes to an error queue | Reasons about it, retries, or escalates |
| Main risk | A wrong answer | Breaks when a system changes | A wrong action on a live system |
What to watch for
- Ask which systems the agent can write to, and insist that every write passes through an approval workflow until the agent has a track record.
- Give each agent its own identity and credentials rather than a shared service account, so the audit log shows who did what.
- Measure task completion and exception rates on real work before expanding scope; a demo on clean data says little about messy purchase orders.
- Check how the agent behaves when a tool fails or returns nothing, since silent retries can create duplicate records.
- Keep the first deployment narrow, with one goal, a handful of tools, and a person reviewing the output.