AI Agent
A system in which an LLM plans and executes multi-step tasks by calling tools, observing results and deciding what to do next, with defined limits.
Definition
An AI agent wraps an LLM in a loop. Given a goal, the model decides on a next step, which is often a tool call (search a database, read a file, call an API). The application runs the tool, feeds the result back, and the model decides again. The loop ends when the model reports completion, a step budget runs out or a rule stops it. Memory, a set of permitted tools and a stopping policy are the other essential parts.
Agents range from narrow (one workflow, three tools, five steps maximum) to broad (a coding or research assistant with dozens of tools). In business settings the narrow kind is far more common and far more reliable. The term "agentic workflow" usually means a fixed sequence of steps where the model handles judgement at each one.
Example: a finance agent receives "reconcile this week's Stripe payouts with the ledger". It pulls payouts via an API, queries the ledger, matches rows, drafts a list of exceptions and asks a human to approve two adjustments before posting them.
Why it matters for your business
- Agents complete work, not just conversations: reconciliations, triage, data entry across systems, research.
- Every tool is a privilege. Least-privilege access and approval gates decide whether an agent is safe.
- Reliability drops as steps multiply; short, well-defined workflows succeed where open-ended ones fail.
- Tracing, evals and cost limits are required because each run can take a different path.
How Teksmobile uses it
Our AI agents and automation service builds narrow, well-instrumented agents on Claude, OpenAI and open models, connected to client systems through MCP servers. We harden them with the controls described in our prompt injection defence work.
Related terms
Part of the Teksmobile AI Glossary. For longer reads, see our guides.
Frequently asked questions
How is an agent different from a chatbot?
A chatbot answers; an agent acts. The agent runs a loop of planning, tool calls and observation until the task is done or a limit is reached. That autonomy is what makes it useful and what makes it risky.
Do we need a framework to build agents?
Not necessarily. Provider SDKs with tool use, plus MCP for connecting systems, cover most cases. Frameworks help with orchestration and tracing but add their own abstractions to maintain.
What should an agent never do without a human?
Anything irreversible or costly: payments, deletions, external communications to customers, contract changes. Put those behind explicit approval steps. See agentic workflows 101.