An AI agent selects and carries out steps towards a goal, using results from one action to decide what to do next. Its tools might let it search sources, query a database or change a record. Software agents predate generative AI; many current systems use a large language model to interpret requests and direct the work.
‘Agentic’ is used for varying degrees of autonomy. A predefined workflow follows a route set by people; an agent can choose or adapt its steps within the tools and permissions it has. A clear, repeatable task may be better served by a fixed workflow. A run may end when the goal is met, a configured limit is reached, it is cancelled or the system cannot proceed without help.
Why it matters for content
A weekly review can use a simple date filter to identify overdue guidance. An agent could then examine each page, choose which approved sources to check and change its next step when it finds a conflict, for example by requesting an editor’s decision rather than drafting through the uncertainty. The schedule is fixed; the route through the evidence can vary.
Connecting these steps can reduce the coordination needed between finding a problem and preparing a useful response. The benefit is greater where findings require different follow-up actions, rather than the same sequence every time. A drafting tool produces material for someone to act on; an agent can move the work towards a reviewable result.
Preparing a private draft and updating a public policy need different permissions and approval rules. Teams should specify what the agent may change, which sources it may use and which actions require approval. Tool permissions should enforce those boundaries where possible; instructions alone don’t prevent an unauthorised action.
How agentic AI changes the picture
Agentic AI can carry work from a brief through research, drafting, checks and preparation for publication. Instead of operating a separate tool at each stage and moving material between them, a team can delegate some of that coordination. An agent can respond to findings, repeat a failed step or send an unresolved issue for review, within a process that still has fixed rules and approval points.
Multi-agent setups divide work between agents with different roles, instructions or tools. One might gather source material while another drafts and a third checks the result against editorial criteria. Their hand-offs need to carry evidence and unresolved questions, not just finished text. A reviewer agent may share the drafting agent’s blind spots, so another pass isn’t automatically an independent check. Adding agents can also increase coordination and cost; the division of work should serve the task.
For teams using agents, content work includes designing and maintaining that process. Editorial standards need to become usable instructions and evaluation criteria, supported by appropriate sources. Content people need to test what the system produces, investigate failures and improve how the work happens, alongside technical colleagues. The skills involved extend beyond writing and editing individual pieces to shaping the systems that produce them.
Questions people ask
- How do agents fit into content workflows?
An agent can handle one stage or several connected stages. A workflow might have fixed review and approval points while allowing an agent to choose which source to consult or which check to run next. Some steps remain predictable; others adapt to what the agent finds.
- Is an AI agent just a chatbot?
A chatbot provides a conversational interface; behind it may be a fixed workflow or an agent. An agent can also work without a chat interface. Using a tool alone doesn’t show that the system chooses its own route.
- Can an agent publish content?
An agent can publish if its tools and permissions allow it. Whether it should depends on the risk, the permissions and the review process. Filing a private draft summary is one thing; updating a public policy page is quite another.
- How much autonomy should an agent have?
Match autonomy to the task and the consequences of error. An agent may select research steps independently but require approval before sending a message or publishing a change. Test its performance on realistic cases, inspect failures and keep responsibility with a named owner.
Learn more
- Agentic AI, explained – MIT Sloan on systems that perceive, reason and act across multi-step tasks, with examples from banking and retail and the risks organisations need to manage.
- Building Effective AI Agents – Anthropic’s 2024 distinction between predefined workflows and agents that choose their steps, with a preference for the simplest system that works; the page notes that its tooling has since changed.
- I think “agent” may finally have a widely enough agreed upon definition to be useful jargon now – Simon Willison’s September 2025 working definition, “An LLM agent runs tools in a loop to achieve a goal”, and where other definitions cause confusion.