What an AI agent is, in one paragraph

An AI agent is a large language model given tools and a goal instead of a single question. It can read files, run code, check a calendar, post to your CMS, and keep working through multi-step workflows until the job is done or stuck. That is the whole trick. Textbooks will offer you a taxonomy on top, simple reflex agents acting on condition-action rules, model-based reflex agents tracking an internal state, utility-based agents scoring outcomes with a utility function, learning agents improving from feedback. Useful for an exam. In production the questions that matter are plainer: what tools does it hold, what data can it reach through your APIs, and who checks its work.

Agents versus RPA and the older ai tools

The nearest ancestor is RPA, robotic process automation, which has been around for fifteen years and is well established in enterprise systems. RPA bots follow predefined scripts and need manual reconfiguration every time the process shifts. Agents are different in kind: they read context, adapt to new situations, and carry judgement through a task. That flexibility is the value and the risk, which is why the rest of this post is mostly about control.

What we actually deploy

Client work at Calibre falls into three patterns, whether the client is a small business or a marketing team inside a larger organisation.

Scheduled workers

Agents that run on a clock without anyone watching. Ours check search indexing daily, reconcile ad spend across platforms each morning, and flag site errors before a client sees them. They do not tire: an agent runs continuously, overnight and on weekends, reading data streams and surfacing what matters in real time. For clients, the same pattern covers stock alerts, review monitoring, and weekly performance summaries written in plain English for non-technical teams. If a task is repetitive, has a clear definition of done, and hurts when a human forgets it, it belongs here.

On-demand specialists

Specialised agents that wait until a person hands them a job: turn this brief into a shot list, pull every brand mention this week, draft replies to customer support tickets in your voice for a human to approve. The person stays in the loop, so the failure mode is cheap. A bad draft costs a review; a bad autonomous action costs a cleanup. Most sales teams and marketing teams should start here.

Long-running autonomous jobs

The newest category and the one we are most careful with. For work that takes hours rather than minutes, we build on Claude Managed Agents: they run asynchronously in secure, managed infrastructure (or your own, where data residency or a regulated industry requires it), reading files, running code and carrying multi-step workflows through to completion without a human holding the thread. We use this for large audits and migrations. We do not use it for anything that messages your customers. Not yet.

The numbers vendors quote, and the one they skip

The pitch decks all carry the same statistics: 74% of executives report ROI from AI agents within the first year, 39% of organisations say productivity gains doubled with AI tools, 62% were already experimenting with agents in 2025, and Gartner expects 40% of enterprise applications to feature AI agents by the end of 2026. The number that matters more: Gartner also predicts over 40% of agentic AI projects will be cancelled by 2027, on cost, unclear value or poor risk control. Both sets of numbers are true. The difference between the companies in the first group and the second is almost never the model. It is ownership, verification and scope.

Where agents fail

Three failure patterns show up over and over, and vendors will not tell you about them.

They fail silently. A cron job that breaks throws an error. An agent that drifts produces plausible output that is quietly wrong. Every agent we ship has a verification step, a second check, sometimes multi-agent orchestration where one agent's only job is to grade another's work, that has to pass before anything counts as done.

They inherit your mess. An agent working across a tidy Notion, a chaotic Drive and three abandoned Slack channels will be exactly as confused as a new hire on day one, except it will not ask. The unglamorous truth is that most "agent projects" are 60% cleaning up the systems the agent has to touch.

They get overbought. Half the briefs we see for "an AI agent" describe a three-step automation with no judgement in it. If nothing in the task requires reading, deciding or writing, you do not need an agent, and we will say so in the first call. The same goes for "AI employees": renaming an agent does not change what it can be trusted with.

Platforms, builders and what this costs

You will meet a wall of AI agent platforms and no-code agent builders, each with a free plan, a team plan and custom enterprise pricing, all promising custom AI agents without the learning curve. Some are competent. But the platform is rarely the decision that matters. A useful scheduled worker is days of work to build, not months, if your data is reachable. The long tail is maintenance: models change, APIs change, your business changes. We structure agent work as build plus a light monthly retainer for exactly that reason, and we would be suspicious of anyone selling a build with no plan for month three.

Whatever you choose, write the governance down before the build: which systems the agent can touch, under whose account, with what audit trail. Security and accountability are design inputs, not a cleanup task. Human expertise does not leave the loop; it moves up a level, from doing the task to owning the system.

Where to start

Pick one measurable use case: a task that is repetitive, well-defined and annoying. Give an agent that job, on-demand first, scheduled once you trust it. The vendor benchmark says agents save 40 to 60 minutes per person per day; treat that as a hypothesis to test against your own numbers, not a promise. Then expand. Teams that start with "an AI strategy" produce decks. Teams that start with one working agent produce momentum.

If you want the fuller picture of how we automate studio and brand workflows, read Intro to AI automation. If you are weighing paid placement inside AI assistants while you are at it, Running ChatGPT Ads covers what the first campaigns taught us. And if you would rather hand the problem to someone who does this daily, that work lives under Intelligence.