Skip to content
Mia Automation
NLEN Sign in Put Mia to workGet started

What Is Agentic AI? A Plain Definition and the Guardrails That Make It Usable

Agentic AI is AI that is given a goal and carries out the work itself: it plans steps, takes actions with tools such as files, email or a browser, and checks the outcome before it moves on. Generative AI produces content and stops; agentic AI acts on it. That shift makes the limits around the agent as important as the model itself. Below: a definition without the buzzwords, a test that shows the difference, and the three limits we enforce on every agent we run.

AI agents and agentic workUpdated 8 min read

Tested by us on 11 November 2026.

Agentic AI, defined

"Agentic" describes behavior, not a type of model. A system is agentic when a language model decides the next step, takes it with a tool, looks at the result and decides again, until the goal is met or a limit stops it. The program doing that is an AI agent. Agentic AI is that agent together with the tools it can reach, the rules it runs under and the checks on its work. The invoice job below shows all four in one run.

Anthropic draws the line between workflows, where code fixes the path and a model fills in steps, and agents, where the model directs its own process and tool use. Its advice is to start with the simplest version that works and only add autonomy when it pays for itself. That matches what we see: most business automation should stay a workflow, with an agent in the few steps that need judgment.

Animated diagram of three agent actions passing a hook: editing a file runs, sending an email waits for a yes, and changing its own rules is refused.
What makes agentic AI safe to run: every action passes a hook that runs outside the model.

Agentic AI vs. generative AI, tested on one task

We gave the same job to a model twice: compare six invoices against a bank export, find what's still unpaid and draft reminders for anything overdue. The data had two traps: a partial payment and a payment reference missing its dash.

One run each on November 11, 2026, with Claude Code in non-interactive mode; the chat run had all tools switched off.
As a chat (generative)As an agent (agentic)
What we didPasted both files into one messageGave one instruction in the folder
Answer correct?Yes, both traps caughtYes, both traps caught
Manual steps left for a person7: open, copy and paste two files, then save two reminders by hand0: two reminder files written, nothing sent
Extra judgmentNo remark about the export's date rangeFlagged that the bank export ended November 5, so a recent payment might be missing

On this task the reasoning was the same and the difference was who did the hands-on work. On a task that needs live data, the chat couldn't do it at all.

  1. Instruction

    One sentence: find which invoices are unpaid or partly paid, draft a reminder for each overdue one in a reminders folder, send nothing, and finish with an overview.

  2. Limits set outside the prompt

    Allowed: reading and writing files, listing and creating folders. Everything else, including other shell commands, needs permission it didn't have.

  3. Read and match

    It read both files and matched payments to invoices, including the payment whose reference lacked a dash and the customer who paid half.

  4. A gate fires

    It tried to create both reminder files with one shell command. The permission settings refused it. Its next attempt used the file-writing tool, which was allowed.

  5. Deliver and stop

    Two reminder drafts in the folder, an overview of €3,327.50 open of which €3,025 overdue, and a warning that the bank export ended on November 5. Nothing was sent; a person reviews the drafts.

A second test shows where the gap is bigger than a few clicks. We asked both modes to check which of six links in a document still work. The chat couldn't: without tools it has no way to request a URL, so it said so and handed back a script. The agent ran the requests and reported four working links and two dead ones in 18 seconds. Both tests ran once per mode on November 11, 2026, so treat them as examples rather than a benchmark; the full write-up is in agentic AI vs. generative AI.

Is ChatGPT agentic AI?

A normal chat with ChatGPT or Claude is generative: you ask, it answers, you do the rest. Both now ship agent features that browse, click or edit files for you, and coding agents like Codex and Claude Code are agentic by design. The useful test is how you use it. If you hand it a goal and tools and it chains steps without you prompting each one, it's acting agentically.

The three guardrails we put on every agent

An agent that acts needs limits it can't move itself. Ours live in code, in hooks that run before each action, not in the prompt. A rule in a prompt is a request; a hook is a limit.

Guardrail 1: What the agent may do on its own

Each topic gets one of three settings: do it, do it and tell us, or ask first. Topics include spending money, sending email, posting to social media and shipping to customers. Money has a daily cap. A few things are never allowed under any setting, such as the agent changing its own rules or storing card numbers. Only the owner can change the settings; the agent can't write to that file.

Bar chart of 1,402 log entries: 1,128 actions done on their own, 37 done and reported, 63 after approval, 99 blocked, 15 refused and 60 logged only.
All 1,402 entries our permission hook recorded from October 8 through November 10, 2026 (our own count). "After approval" includes actions the owner let through once; "logged only" covers sessions outside the permission scheme.

Of the 99 blocked actions, 72 were attempts by an agent to edit its own rules or the guards around it; the other 27 involved money. If those limits had lived only in the prompt, we'd have been relying on the model's restraint 99 times.

Guardrail 2: Where the agent stops

Every job gets a time limit (six hours by default for us) and a memory cap, so a stuck agent ends by itself. When it needs a human decision, it asks one self-contained question with its recommendation and two to four concrete answers to pick from, then waits.

Guardrail 3: Who checks the work

An agent never gets to declare itself done. Code has to pass tests. Writing has to pass a style gate and a second model that scores it. A screen has to pass a fixed twelve-point review. Work that misses the bar goes back. The agent's last act is a report: what's finished, the evidence, and what's still open.

When an agent gets stuck

Long jobs hit walls: usage limits, server restarts, a question waiting on an answer. Of our 1,221 finished jobs since October 4, 2026, 355 needed more than one run. Starting over would mean paying for the same work twice, so we keep each session and resume it.

  • Usage limit: the resume is scheduled for when the limit resets, or after 10, 30 or 60 minutes if that time is unknown.
  • Waiting on a person: no automatic resume; it continues only with the answer, so it never guesses past a decision it needed.
  • Background work still running: it gets a follow-up turn to finish that work, at most four times, after which the job is marked failed and we're notified.
  • Broken for real: the job ends as failed with a notice. That happened to 5 of the 1,373 jobs we started; of the rest, 1,221 finished, 137 were stopped by hand and 10 were still running on November 11.

Questions

What is agentic AI in simple terms?

AI that's given a goal and does the work itself: it plans steps, uses tools such as files or email, and checks the result. It's a way of running language models, not a separate kind of model.

What is the difference between agentic AI and generative AI?

Generative AI creates something and stops. Agentic AI also carries out the actions. In our test both gave the same correct answer, but the chat left seven manual steps and the agent left none.

What is an example of agentic AI?

An agent that matches your bank export against your invoices, works out what's overdue and drafts a reminder per customer without sending anything. Coding agents that edit a codebase and run its tests are another everyday example.

What are the risks of agentic AI?

Errors compound across steps, costs rise with every action, and an agent can do something you didn't intend. Put permissions in code, give every job a time limit, and require a check before work counts as done.

What happens when an AI agent gets stuck?

With a time limit it stops on its own. If you keep the session, it can pick up where it left off after a limit or outage. If it's waiting on a person, it should resume only with the answer.

Sources