Agentic AI is software in which a large language model (LLM) works toward a goal by deciding its own next steps: it calls tools, observes the results, and loops until the task is done, instead of answering a single prompt and stopping. AI coding agents that read a codebase, edit files and run tests on their own are the most familiar example.
Anthropic's engineering post Building effective agents (December 2024) gives the clearest working definition. Agents "are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." This guide explains the agent loop, how agents differ from chatbots and workflows, where they show up in coding, and how to keep them safe.
What does agentic AI mean?
"Agentic" describes how much the model is in charge of the process. Anthropic groups a range of designs under one term: "we categorize all these variations as agentic systems." Within that, it draws one important line:
- Workflows: "Workflows are systems where LLMs and tools are orchestrated through predefined code paths."
- Agents: systems where the LLM decides the path itself, choosing which tools to call and when to stop.
The building block for both is what Anthropic calls the augmented LLM: "an LLM enhanced with augmentations such as retrieval, tools, and memory." A model that can only write text becomes agentic when it can act through tools and see what happened.
How does the agent loop work?
Every agent runs some version of the same loop:
- Goal. A person or a system gives the agent a task.
- Decide. The model reads the task and everything in its context and chooses the next action.
- Act. It calls a tool: run a command, read a file, query an API, search the web.
- Observe. The tool result comes back into the model's context.
- Repeat or stop. The model decides whether the task is done, needs another step, or needs a human.
Anthropic sums it up: agents "are typically just LLMs using tools based on environmental feedback in a loop." It stresses that the agent must get "ground truth" from the environment at each step, "such as tool call results or code execution," to judge progress, and that it is "common to include stopping conditions (such as a maximum number of iterations) to maintain control."
Three parts make up the loop:
| Part | Role | Example in a coding agent |
|---|---|---|
| Model | Reasons and chooses the next action | A Claude, GPT or Gemini model |
| Tools | Let the model act and observe | Read file, edit file, run shell command, web search |
| Environment | Where actions happen and feedback comes from | Your repo, terminal, test suite, CI |
The mechanism behind step 3 is tool calling: the model returns a structured request to call a named tool, your code runs it, and the result is passed back.
How is agentic AI different from chatbots and workflows?
| Chatbot | Workflow | Agent | |
|---|---|---|---|
| Who decides the steps | The user, one message at a time | The developer, in code | The model, at run time |
| Uses tools | Sometimes, one call at a time | Yes, in a fixed order | Yes, chosen per step |
| Typical length | One turn | A fixed number of steps | Open-ended, until done or stopped |
| Predictability | High | High | Lower |
| Best for | Q&A, drafting | Repeatable pipelines | Tasks where the steps cannot be known in advance |
Anthropic's guidance on when to use agents is specific: "Agents can be used for open-ended problems where it's difficult or impossible to predict the required number of steps, and where you can't hardcode a fixed path." For well-defined tasks, workflows "offer predictability and consistency."
It also warns against reaching for agents too early: "we recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all."
What does agentic AI look like in coding?
Coding is where agents have taken off fastest, because code gives fast, objective feedback: tests pass or fail, builds compile or do not. Anthropic lists "A coding Agent to resolve SWE-bench tasks, which involve edits to many files based on a task description" among its own examples.
Claude Code's How Claude Code works page describes its loop as three phases: "gather context, take action, and verify results." It splits the roles as "models that reason and tools that act," and says "Tools are what make Claude Code agentic. Without tools, Claude can only respond with text."
Asked to "fix the failing tests," the docs say Claude might run the test suite, read the error output, search for the relevant source files, read them, edit them, and run the tests again. No step was scripted in advance; each one depended on the last result.
Other forms you will see:
- Background agents that take an issue, work in a cloud sandbox and open a pull request.
- Subagents that a main agent spawns for a narrow job, each with its own context window.
- Review agents that read a diff and leave comments.
Our agentic coding guide covers how to work with these day to day, and what an AI coding agent is compares the main tools.
What are the risks of agentic AI?
Giving a model the ability to act changes the risk profile. Anthropic is direct: "The autonomous nature of agents means higher costs, and the potential for compounding errors." The main risks:
- Compounding errors. A wrong assumption early can lead to many confident wrong steps.
- Real side effects. Agents can delete files, push code, call paid APIs or change production data.
- Prompt injection. Claude Code's security docs define it as "a technique where an attacker attempts to override or manipulate an AI assistant's instructions by inserting malicious text." A web page, issue or file the agent reads can carry that text.
- Cost and time. Loops that run many steps use many tokens, and each step adds latency.
- Context drift. Long sessions fill the context window, and summaries can drop early instructions.
Which guardrails keep agents safe?
Anthropic recommends "extensive testing in sandboxed environments, along with the appropriate guardrails." In practice, coding agents layer several controls:
- Permissions. Claude Code's docs describe modes ranging from asking before edits and commands to auto modes where a classifier blocks risky actions.
- Sandboxes. Claude Code can sandbox bash commands "with filesystem and network isolation," and its cloud sessions run in isolated VMs.
- Checkpoints. Claude Code snapshots files before editing so you can rewind, though the docs note actions on remote systems "can't be checkpointed."
- Stopping conditions. Cap iterations, time or spend so a stuck loop ends.
- Human checkpoints. Review plans before execution and diffs before merge.
- Untrusted content hygiene. The security docs advise you to "Avoid piping untrusted content directly to Claude" and to use VMs when interacting with external web services.
Anthropic's three core principles for agents apply to any vendor: "Maintain simplicity in your agent's design," "Prioritize transparency by explicitly showing the agent's planning steps," and "Carefully craft your agent-computer interface (ACI) through thorough tool documentation and testing."
How do you build an agent?
You can build the loop yourself with any model API that supports tool calling, or use a framework. The OpenAI Agents SDK docs describe agents as "LLMs equipped with instructions and tools," with handoffs that "allow agents to delegate to other agents for specific tasks" and guardrails that "enable validation of agent inputs and outputs." Anthropic, by contrast, notes that the most successful implementations it saw used "simple, composable patterns" rather than complex frameworks.
Either way, the quality of the instructions matters as much as the loop. A clear task brief, a definition of done and a list of things the agent must not touch go a long way. Kipt keeps those briefs and Agent Skills as Markdown in a private GitHub repo you own, and can install your skills into Claude Code or Codex with one command, so the instructions that work travel with you.
Compare options in our AI agent frameworks guide.
What should you read next?
Start with what an LLM is if the model side is new, then tool calling to see how agents act. For hands-on practice, read agentic coding and AI agent frameworks.