Resources, AI agents
What are AI agents? A practical guide for companies
An AI agent is software that uses a language model to work towards a goal. It plans steps, uses tools and checks the results until the task is done. This guide explains what that means in practice, and where agents help or fail.
Updated 20 November 2026. 9 minute read.
In short
- An AI agent is a system in which a language model directs its own steps and tool use in a loop to complete a task, instead of following a fixed script or giving a single answer.
- Chatbots answer questions, copilots assist an employee inside one app, and workflow automation and RPA follow predefined steps. An agent decides the next step itself, within the limits it is given.
- Every agent has five parts: a model, tools, memory, instructions and a loop that repeats until the task is done or a limit is reached.
- Agents bring specific risks: prompt injection, excessive permissions, compounding errors and higher costs. Set the level of autonomy per action and keep an employee in charge.
- Start with a frequent, checkable task that resists simple rules, can be undone and has a named owner.
What an AI agent is
There is no single legal or technical definition of an AI agent. Anthropic's engineering guidance draws a useful line. In a workflow, language models and tools follow code paths that a developer set in advance. In an agent, the model dynamically directs its own process and tool use. OpenAI's guide to building agents describes them as systems that independently accomplish tasks on your behalf.
In practice, an agent receives a goal, plans steps and calls tools such as search, email or a database. It looks at each result and decides what to do next. Anthropic describes agents as typically just language models using tools based on feedback from their environment, in a loop. The loop is what sets an agent apart. It keeps working until the task is done, it needs input, or it reaches a limit.
The EU AI Act does not use the word agent. An agent will usually meet the definition of an AI system in Article 3(1): a machine-based system designed to operate with varying levels of autonomy that infers from its input how to generate outputs such as content, recommendations or decisions. Your obligations depend on your role and the use case, not on the label.
A quick test: if the software decides which step comes next, it is acting as an agent. If a developer fixed every step in advance, it is a workflow, even when a language model writes some of the text.
How agents differ from chatbots, copilots, workflows and RPA
The word agent is now used for many older products. Gartner calls this agent washing: rebranding AI assistants, chatbots or robotic process automation (RPA) tools as agentic AI without real agentic capability. In June 2025, Gartner estimated that only about 130 of the thousands of vendors claiming agentic features offered genuine ones. The differences below help you tell them apart.
- Chatbot: answers questions in a conversation. It responds to each message and usually takes no action outside the chat.
- Copilot: an assistant built into an application that helps an employee with the task in front of them. The employee prompts it, reviews the suggestion and decides what to do.
- Workflow automation: a fixed sequence of steps started by an event, such as a new form entry creating a record in a CRM. A language model may handle one step, but the path is set in advance.
- Robotic process automation (RPA): software robots that repeat rule-based actions in applications, such as copying data between systems. It is dependable while screens and rules stay the same, and breaks when they change.
- AI agent: given a goal, it chooses the steps and tools itself, adapts to what it finds and can work across several systems. It can also use a workflow or an RPA robot as one of its tools.
These categories overlap and are often combined. Anthropic recommends the simplest solution that works: many tasks need only one well-designed model call or a fixed workflow. Agents earn their place when the number of steps cannot be predicted in advance and the task needs judgement.
The five components of an agent
Every agent combines five parts, however it is built. OpenAI names the first three as the foundations: model, tools and instructions. Anthropic describes the basic building block as a language model with retrieval, tools and memory. The loop turns that building block into an agent.
- Model: the language model that reasons, plans and writes. Larger models handle complex, ambiguous tasks better. Smaller ones are faster and cheaper. Anthropic suggests routing simple requests to smaller models and hard ones to more capable models.
- Tools: the functions the agent can call, such as reading a calendar, searching documents, sending an email or updating a record. Tools give the agent its reach, and so most of its risk. Open standards such as the Model Context Protocol (MCP) let one agent use tools from many providers.
- Memory: what the agent knows beyond the current request. Short-term memory is the context of the current task. Long-term memory sits outside the model, such as notes, past decisions and company documents that the agent retrieves when needed. Anthropic reports that models recall information less accurately as the context fills up, so good agents load only what they need.
- Instructions: the role, goal, rules and limits written for the agent, such as which tool to use for which step and when to stop and ask. OpenAI recommends building them from existing operating procedures, support scripts and policy documents, broken into clear steps.
- Loop: the cycle of plan, act, observe and decide. The loop ends when the goal is met, an employee must approve a step, or a limit on steps, time or cost is reached. OpenAI's guide lists limits on retries and actions as a guardrail.
If you change the model but keep the tools, memory and instructions, you keep most of the agent. Store memory and instructions where they do not depend on one AI provider.
Levels of autonomy
Autonomy is how much an agent does without an employee's involvement. Researchers at the University of Washington proposed five levels in 2025, defined by the role the employee plays:
- Operator: the employee directs and makes the decisions, and the agent carries them out.
- Collaborator: the employee and the agent plan, delegate and carry out the work together.
- Consultant: the agent leads and asks the employee for expertise or preferences.
- Approver: the agent works alone and involves the employee only in risky or pre-specified situations.
- Observer: the agent works with full autonomy while the employee monitors it.
Autonomy is a design choice, not a property of the model. One agent can work at different levels for different actions. It may draft customer replies as a collaborator and need approval for every refund. Set the level per action, and raise it only when the activity log shows the agent handles that action well.
Common business uses
Agents fit tasks that have resisted rules-based automation. OpenAI's guide names three signs: complex decisions that need judgement, rules that have become hard to maintain, and heavy reliance on unstructured data such as documents and emails. Anthropic points to customer support and coding as areas where agents already work well, because both combine conversation or reasoning with actions, and their results can be checked.
- Customer support: read incoming requests, look up orders, draft replies and pass complex cases to an employee.
- Research and analysis: search sources, compare findings and write a briefing with references.
- Software engineering: write and test code changes, with automated tests as an objective check.
- Sales and marketing: research accounts, draft outreach and content for review, and keep the CRM up to date.
- Operations and finance: match documents, chase missing information and prepare reports for sign-off.
- Internal knowledge: answer employees' questions from company documents and keep notes current.
In each case the agent prepares the work and an employee owns the result. The value comes from time freed for review and decisions. Measure that value instead of assuming it.
Limits and risks
Agents fail in ways that ordinary software does not. In June 2025, Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear business value or inadequate risk controls. Most of these causes can be managed by design.
- Prompt injection: instructions hidden in an email, web page or document can redirect the agent. OWASP ranks prompt injection first in its 2025 Top 10 for LLM applications, and its Top 10 for agentic applications puts agent goal hijack first.
- Excessive agency: an agent with more tools, permissions or autonomy than its task needs does more damage when it gets something wrong. OWASP names excessive functionality, excessive permissions and excessive autonomy as the root causes.
- Compounding errors: one wrong step early in a long task can send the rest off course. Anthropic notes that agents keep state, so errors compound.
- Over-reliance: models can produce output that sounds right but is wrong, and reviewers can come to trust it too much. The EU AI Act calls this automation bias in Article 14.
- Cost: Anthropic measured that agents use about four times as many tokens as chat interactions, and multi-agent systems about fifteen times as many.
- Personal data: agents often read customer or employee data. The GDPR applies to that processing as it does to any other.
Give each agent the least agency its task needs: only the tools, permissions and autonomy required. OWASP makes this its guiding principle for agentic applications.
Choosing the first use case
A good first use case is narrow enough to measure and valuable enough to matter. Score each candidate against six questions:
- Is the task frequent? A task done once a month gives too few examples to learn from.
- Does it resist simple rules? If a fixed workflow would do the job, use a workflow.
- Can the result be checked? A clear definition of done, or an objective test, makes evaluation possible.
- Can mistakes be undone? Start with drafts, internal updates and recommendations, not payments or customer messages sent without review.
- Does it have an owner? Name the employee who approves the agent's work and answers for it.
- Are the data available and allowed? Check that the agent can reach the systems it needs, and that the GDPR and your own policies permit the use.
Then run the agent at a low level of autonomy on real work for a few weeks. Collect the cases where it fails and turn them into an evaluation set. Anthropic suggests that 20 to 50 simple tasks drawn from real failures is a good start. Raise autonomy one step at a time, and only on evidence.
How Blentera helps
Blentera is a platform where companies run AI agents with employees in charge. Each agent has a role in the org chart. Approval rules can be set per agent, project, app or tool, so autonomy can be set per action. Budgets warn at 80% and pause the agent at the limit. One company memory stays when you switch AI models, and an activity log records what agents do. Connectors link agents to apps and to MCP servers, and routines start on a schedule or from a webhook. Blentera runs its own native agents as well as OpenClaw, Hermes and Cursor Cloud agents, and it is hosted in the EU.
Questions
- A chatbot answers messages in a conversation and usually takes no action outside it. An AI agent works towards a goal: it plans steps, calls tools such as email or a database, checks the results and decides what to do next, in a loop, until the task is done or it needs an employee.
- No. RPA robots repeat fixed, rule-based actions and break when screens or rules change. An AI agent uses a language model to decide its own steps and can handle unstructured input such as emails and documents. An agent can use an RPA robot as one of its tools.
- A model that reasons and writes, tools that let it act, memory that holds context and company knowledge, instructions that set its role and limits, and a loop of planning, acting and checking that runs until the goal is met or a limit is reached.
- The AI Act does not define agents, but an agent usually meets the definition of an AI system in Article 3(1). What you must do depends on your role, usually deployer, and on the use case. Duties such as AI literacy apply to every company that uses AI systems.
- A frequent task that resists simple rules, has a checkable result, can be undone if the agent gets it wrong and has a named owner. Drafting support replies for review or preparing research briefings are typical examples.
Sources
- Building effective agents, Anthropic Engineering
- A practical guide to building agents, OpenAI
- Effective context engineering for AI agents, Anthropic Engineering
- How we built our multi-agent research system, Anthropic Engineering
- Demystifying evals for AI agents, Anthropic Engineering
- Levels of Autonomy for AI Agents, Feng, McDonald and Zhang, arXiv:2506.12469
- Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027, Gartner press release, 25 June 2025
- OWASP Top 10 for LLM Applications 2025: LLM06 Excessive Agency, OWASP GenAI Security Project
- OWASP Top 10 for Large Language Model Applications, OWASP
- OWASP Top 10 for Agentic Applications for 2026, OWASP GenAI Security Project
- Model Context Protocol specification, version 2026-07-28
- Regulation (EU) 2024/1689 (AI Act), consolidated text of 27 July 2026, EUR-Lex
- Regulation (EU) 2016/679 (GDPR), EUR-Lex
Related
This guide is general information, not legal advice. Check how the rules apply to your company with your own advisers.
Get started
Put your first agent to work in minutes.
Three steps, and your first agent is working for your company.
Step 1
Create an account
Sign up with your work email and invite your team. No sales call needed.
Step 2
Add your first agent
Create an agent, give it a role, and connect your own AI key and the apps it may use.
Step 3
Give it a task
Ask in Blentera or your chat app, and follow the work until it is done.