Agentic systems are a new class of AI — they break down goals, choose tools, write code, and complete real work without being hand-held at every step.
Capabilities
Unlike a chatbot, an agentic system doesn't just answer — it acts, reasoning over multiple steps to finish a job.
Turn a fuzzy goal into a concrete sequence of steps, then re-plan as new information arrives.
Call APIs, browse the web, query databases, run code — chained together to get real outcomes.
Run for minutes or hours on long-horizon tasks: research, migrations, ops work, content pipelines.
Work alongside humans and other agents — handing off, asking for review, reporting progress.
Detect failures, read error output, and try a different approach without starting over.
Cite sources, log every action, and keep humans firmly in the loop on consequential steps.
In the wild
Claude can autonomously navigate codebases, edit files across a project, run commands, and verify its own work — the same loop that powers Claude Cowork, where teammates hand work to Claude and review pull requests it produces.
The OpenAI Agents SDK and ChatGPT's agent mode let GPT models plan, call tools, hand off to specialised sub-agents, and complete tasks end-to-end — booking, researching, building, summarising — with traces you can audit.
How it works
A user (or another system) gives the agent a task in plain language.
It decomposes the goal, picks tools, takes actions, and observes results.
It delivers finished work — a PR, a report, a booking — with a full trace.
AgenticNZ tracks how autonomous AI is being put to work — and helps New Zealand teams figure out where it actually fits.