Chatbots answer questions. Agents do work. The leap from "AI that talks" to "AI that acts" is the most important shift in software since the cloud — and it changes how products are built and how they must be secured.
What "agentic" actually means
An agentic system uses a language model as a reasoning engine that can plan a multi-step task, use tools (call APIs, run code, search, query a database), observe the result, and decide the next step — looping until the goal is met. Instead of returning a paragraph, it books the meeting, files the ticket, or reconciles the data.
The building blocks
- Reasoning + planning — breaking a goal into steps.
- Tools / function calling — the hands that let it act in the real world.
- Memory — short-term context plus longer-term state across runs.
- Orchestration — often multiple specialised agents coordinating on a job.
Where agents are useful today
The sweet spot is repetitive, multi-step knowledge work with clear success criteria: customer-support resolution, data entry and reconciliation, research and summarisation, code migration and testing, and internal "do this for me" workflows. Used well, an agent compresses a 30-minute manual process into seconds.
The new security surface
An agent that can act is also an agent that can be manipulated into acting. The risks are real and new:
- Prompt injection — malicious instructions hidden in a webpage, email or document the agent reads, hijacking its behaviour.
- Excessive permissions — an agent with broad API/database access becomes a high-value target.
- Data exfiltration — tricking the agent into sending sensitive data outward.
- Unbounded actions — loops that spend money or make irreversible changes.
How to deploy agents safely
Treat the agent like an untrusted junior employee with powerful tools. Apply least privilege to every tool, keep a human in the loop for irreversible or high-value actions, sandbox code execution, validate and constrain tool inputs/outputs, set hard limits (budget, rate, scope), and log everything for audit. This is exactly the offensive-meets-defensive mindset agentic systems demand.
We build and secure agentic features — least-privilege tools, prompt-injection defences and audit trails baked in. Read our whitepaper on the human edge in AI security, or talk to us.