What Is an AI Agent?

An AI agent uses a model to choose actions toward a goal, act with tools, and learn from the results. The key difference from a fixed workflow: it can adapt its next step.

Is it an agent, or just called one?

Check one product mode at a time. Choose a preset, then adjust its behavior below.

The code or a person decides the path in advance, including its branches.

For example: search, read files, run code or update a record.

It decides which step or tool to use, rather than only filling in a fixed step.

A tool result can change what it tries next, before the task ends.

Saved facts or preferences, beyond the context of the current conversation.

Approval changes the level only for an adaptive agent. It does not turn a fixed workflow into one.

Chatbot

Educational autonomy: 0/5 · Text-only chat

  • It has no external tools to act on the world. Planning text alone is still a response.
  • No cross-session memory is described. An agent can still work using only the context of its current task.

A teaching guide, not a standard, benchmark, certification or measure of intelligence. Definitions overlap. More autonomy does not mean better or safer.

Your answers stay in this tab. No account, model call or saved profile.

What do levels 0–5 mean here?
  1. Text-only chat.
  2. Fixed workflow.
  3. Tool-assisted help.
  4. Supervised agent.
  5. Exception-gated agent.
  6. Bounded delegation: delegated work is still bounded, with stop conditions and human responsibility.

How an agent works: a loop, with an exit

Agent action and feedback loopGoal and limits lead to Choose, Act, and Observe. After checking the result, continue back to Choose or stop and ask a person when done, blocked or at a limit.Goal + limitsChoose an actionAct with a toolObserve the resultDone, blocked or at a limit?Stop / ask a personNoYesRepeat within limits

The observation changes the next choice

An agent might run a test, read the failure and inspect a different file. It can choose a new action from the result, rather than repeat the same preset steps.

Stop when the goal is complete, a budget is reached or progress is blocked. Ask a person when the next action needs permission or the evidence is insufficient.

Observation feeds the next choice. A failed test might lead a coding agent to inspect a different file. A fixed retry rule, by contrast, can loop without giving the model control of the path. Neither pattern needs unlimited execution.

Agent vs chatbot vs workflow vs assistant

A practical distinction, not a universal taxonomy. “Assistant” is often used as a broad product name.
PatternWho chooses the path?Tools and feedbackTypical example
ChatbotUsually the user, turn by turnMay retrieve context; the described mode does not act through toolsAnswer a question about a policy
WorkflowPredefined code and branchesCan use LLMs, tools, loops and checksExtract an invoice, validate it, route it for approval
AssistantMostly the user or a single requested stepCan call tools without a sustained adaptive loopLook up a meeting time on request
AI agentThe model chooses within constraintsAction results inform what it tries nextInvestigate a bug, edit code, run tests and revise

This distinction follows the architectural idea that workflows use predefined paths while agents direct their own tool use. See Anthropic’s explanation and LangGraph’s workflow and agent patterns. The checker’s autonomy ladder is our own educational shorthand.

40 documented examples, with the machinery visible

Real products and runnable sample implementations, checked November 16, 2026. Samples are learning material, not claims of production readiness. Some are fixed workflows or tool-assisted systems: the point is to inspect how they work, not accept every “agent” label.

Model versions and permissions vary. “Context” means information available for a task; it does not promise long-term memory. Each source describes the listed product or sample, not an independent performance test.

40 of 40 examples

ADK skills tutorCoding · Sample
Model
Gemini (recipe configuration)
Tools
Skill discovery and resource loading
Memory
Session context; reusable skills on disk
Loop
Select and load skills as the task needs them
Read the sample source
Brand-aligned presentationsSales · Sample
Model
Gemini (recipe configuration)
Tools
Search, internal retrieval, PowerPoint rendering
Memory
Research anchors in session state
Loop
Research, review, outline and render with user checkpoints
Read the sample source
Brand search optimizationSales · Sample
Model
Gemini (recipe configuration)
Tools
BigQuery and browser search
Memory
Catalog and report context
Loop
Find keywords, inspect results and critique title suggestions
Read the sample source
Code security fixerCoding · Sample
Model
Gemini (recipe configuration)
Tools
Bandit and deterministic functional/exploit tests
Memory
Findings and rejected-patch feedback
Loop
Fixed graph: propose, validate and retry within a budget; model itself has no tools
Read the sample source
SQL rewrite optimizerOperations · Sample
Model
Gemini (recipe configuration)
Tools
SQL execution and result/scan-cost verification
Memory
Verified candidates in workflow state
Loop
Fixed propose/verify/select loop with a maximum evaluation count
Read the sample source
Financial advisor sampleResearch · Sample
Model
Gemini (recipe configuration)
Tools
Google Search and specialist agents
Memory
Research passed between specialists
Loop
Research, outline strategies, plan and evaluate risk; educational, no trade execution
Read the sample source
High-volume document analyzerResearch · Sample
Model
Gemini (recipe configuration)
Tools
Authenticated document retrieval and batch analysis
Memory
Batch cursor in session state
Loop
Read batches and stop early when an answer is found
Read the sample source
LLM auditorResearch · Sample
Model
Gemini (recipe configuration)
Tools
Search and claim verification
Memory
Claims and audit context
Loop
Extract claims, check sources, optionally revise the response
Read the sample source
Hybrid document RAGResearch · Sample
Model
Gemini (recipe configuration)
Tools
Document conversion, vector and keyword search
Memory
External document index
Loop
Retrieve evidence and answer; retrieval by itself does not prove autonomy
Read the sample source
On-brand media generatorSales · Sample
Model
Gemini (recipe configuration)
Tools
Image generation, policy lookup and scoring
Memory
Cloud Storage intermediates and scores
Loop
Generate, judge and retry until a threshold or iteration limit
Read the sample source
Retail product searchSales · Sample
Model
Gemini (recipe configuration)
Tools
BigQuery and semantic catalog search
Memory
External catalog; long-term personal memory not established
Loop
Search the catalog and return matching products
Read the sample source
Retail virtual try-onSales · Sample
Model
Gemini (recipe configuration)
Tools
Image try-on and video generation
Memory
Input images and generated artifacts
Loop
Select an image or video tool and return the result
Read the sample source
Small business loan sampleOperations · Sample
Model
Gemini (recipe configuration)
Tools
Document extraction and specialist agent tools
Memory
Firestore repair/resume checkpoints
Loop
Sequential processing with validation and human approval; a workflow example
Read the sample source
Shopping UI agentSales · Sample
Model
Gemini (recipe configuration)
Tools
Product ranking and staged UI widgets
Memory
Widget payloads and preferences in session state
Loop
Stage changes, then flush the resulting interface at turn end
Read the sample source
Voice agent for coding threadsPersonal · Sample
Model
Gemini (recipe configuration)
Tools
Coding threads and connected app tools
Memory
Multi-turn thread context
Loop
Delegate work, announce progress and pause for gated actions
Read the sample source
Ambient expense workflowOperations · Sample
Model
Gemini (recipe configuration)
Tools
Pub/Sub events, risk analysis and approval UI
Memory
Workflow state while awaiting approval
Loop
Fixed amount-based routing; high-value expenses require a manager
Read the sample source
Ambient quality investigatorOperations · Sample
Model
Gemini (recipe configuration)
Tools
BigQuery telemetry and evaluation tools
Memory
Recurring insights with occurrence history
Loop
Investigate traces on demand or schedule and track defects
Read the sample source
Cross-session memory assistantPersonal · Sample
Model
Gemini (recipe configuration)
Tools
Memory Bank recall and extraction
Memory
Persistent facts and preferences across sessions
Loop
Recall, respond and save relevant memories; memory is not itself agency
Read the sample source
ADK deep searchResearch · Sample
Model
Gemini (recipe configuration)
Tools
Web search and report synthesis
Memory
Research context and findings
Loop
Approve a plan, search, reflect on gaps and repeat
Read the sample source
Commerce media studioSales · Sample
Model
Gemini (recipe configuration)
Tools
Photo/video try-on, product spins and optional quality checks
Memory
Studio configuration and output artifacts
Loop
Configure from conversation and request approval before generation
Read the sample source
Long-horizon harnessPersonal · Sample
Model
Gemini (recipe configuration)
Tools
Sandbox, subagents and guarded tools
Memory
Memory Bank plus a persistent per-user sandbox
Loop
Work across turns with tools, feedback and explicit guardrails
Read the sample source
Google Drive readerOperations · Sample
Model
Gemini (recipe configuration)
Tools
OAuth-authorized Drive reading
Memory
Session context; persistent memory not established
Loop
Resolve user authorization, read permitted files and respond
Read the sample source
Search-grounded document assistantCustomer support · Sample
Model
Gemini (recipe configuration)
Tools
Agent Platform Search retrieval
Memory
Indexed documents; not personal memory
Loop
Retrieve source documents and answer a question
Read the sample source
Vector-search document assistantCustomer support · Sample
Model
Gemini (recipe configuration)
Tools
Vector Search retrieval
Memory
Document embeddings in an external collection
Loop
Retrieve matching chunks and respond; an example of an agent component
Read the sample source
Claude CodeCoding · Product
Model
Claude; selected version varies
Tools
File editing, shell and connected tools
Memory
Project instructions and session context
Loop
Inspect code, edit, run checks and adapt to results
Read the product source
Cursor AgentCoding · Product
Model
Selected model; multiple providers
Tools
Code search, edits and terminal
Memory
Project rules and conversation context
Loop
Explore, change files and use results to continue
Read the product source
GitHub Copilot in the cloudCoding · Product
Model
Selected/available model varies
Tools
Repository changes and code review
Memory
Repository and session context
Loop
Plan, implement and review work within a session
Read the product source
Replit AgentCoding · Product
Model
Model selection depends on mode
Tools
Project editing and app testing
Memory
Project and conversation context
Loop
Build from instructions, test and refine the application
Read the product source
DevinCoding · Product
Model
Vendor-managed; exact model not established here
Tools
Code environment, terminal and browser
Memory
Repository knowledge and task context
Loop
Investigate, implement and test changes for review
Read the product source
Sierra customer serviceCustomer support · Product
Model
Vendor-managed; exact model not established here
Tools
Connections to CRM and order systems
Memory
Customer/business context; persistence depends on setup
Loop
Converse and take configured actions to resolve customer tasks
Read the product source
Gemini CLICoding · Product
Model
Gemini; selected version varies
Tools
Files, shell, search and MCP tools
Memory
Context files and session history
Loop
Choose tools and iterate on a terminal task
Read the product source
Amazon Q DeveloperCoding · Product
Model
Vendor-managed; exact model not established here
Tools
Code assistance and development tools
Memory
Project/task context; persistence depends on setup
Loop
Plan and carry out development changes with review
Read the product source
Gemini Deep ResearchResearch · Product
Model
Gemini; version changes over time
Tools
Web research and report generation
Memory
Research task context; long-term memory not asserted
Loop
Form a research plan, gather information and write a report
Read the product source
Zendesk AI agentsCustomer support · Product
Model
Vendor-managed; exact model not established here
Tools
Business knowledge and connected-system actions
Memory
Business and conversation context
Loop
Handle service requests and escalate under configured rules
Read the product source
ManusPersonal · Product
Model
Vendor-managed; exact model not established here
Tools
Browser, software tools and connected apps
Memory
Persistent sandbox files and task context
Loop
Plan and execute multi-step tasks in a virtual computer
Read the product source
Browser Use agentOperations · Library example
Model
Configurable model/provider
Tools
Browser navigation and interaction
Memory
Task history; persistent memory depends on configuration
Loop
Observe pages, choose actions and continue toward a task
Read the library example source
Stagehand browser agentOperations · Library example
Model
Configurable model/provider
Tools
Browser action, extraction and agent APIs
Memory
Browser/task context; persistence is application-defined
Loop
Agent mode can execute multi-step browser work; individual act calls are building blocks
Read the library example source
Open Deep ResearchResearch · Sample
Model
Configurable models by role
Tools
Search providers and optional MCP tools
Memory
Research state and intermediate findings
Loop
Research subtopics and synthesize a report
Read the sample source
LangChain email assistantPersonal · Sample
Model
Configurable model
Tools
Email tools with human-in-the-loop controls
Memory
Demonstrates learned preferences and memory
Loop
Triage and draft, asking for human input where configured
Read the sample source
SWE-agent issue fixerCoding · Sample
Model
Configurable language model
Tools
Repository environment and shell tools
Memory
Issue and action history
Loop
Inspect an issue, edit code and iterate on environment feedback
Read the sample source

Build the smallest system that does the job

  1. Pick a narrow goal and a way to check success. A fixed workflow is often enough when the steps are already known.
  2. If the path must adapt, give the model a few well-defined tools. Validate their inputs and return clear observations.
  3. Set permissions, time and spending budgets, approval points and stop conditions before connecting real accounts.
  4. Keep untrusted retrieved content separate from instructions. Test failure cases and inspect action logs before expanding access.

Choose a workflow pattern or try the Claude Agent SDK starter when you are ready to implement.

How to use the checker

  1. 1. Choose a concrete mode

    Describe one behavior, such as text-only chat, an invoice workflow or a coding agent. Use a preset to get started.

  2. 2. Describe its decisions

    Select whether the path is fixed, which tools it can use, whether the model chooses actions and whether results affect the next step.

  3. 3. Separate memory and permission

    Choose whether it remembers across sessions and who approves its external actions. These are separate design choices.

  4. 4. Read the reasons

    Use the explanation and documented examples to challenge the label. The educational autonomy level is not a safety rating.

AI agent questions, answered

What is an AI agent?

An AI agent is a system that uses a model to choose actions toward a goal, uses tools to take those actions, and observes results to decide what to do next. Definitions vary: this guide distinguishes that adaptive loop from a fixed workflow.

Is ChatGPT an AI agent?

A product name alone does not answer this. A text-only chat is a chatbot; a mode that chooses tools and adapts across steps can be agentic. Classify the behavior of the specific mode and permissions you use.

What are the types of AI agents?

One useful architecture distinction is single-agent versus multi-agent systems. Another is reactive versus planning behavior. Autonomy is separate: either architecture can require human approval. There is no single universal list of agent types.

Are AI agents safe?

Safety depends on the task, permissions, data, tools and controls. An agent can make mistakes or follow malicious instructions in retrieved content. Limit access, isolate execution, record actions, set budgets and stop conditions, test realistic failures, and require approval for sensitive actions. A classification score cannot establish safety.

Does an AI agent need memory?

It needs enough context to work on the current task, but it does not need persistent memory across sessions. A chatbot can have long-term memory without being an agent.

How do I build an AI agent?

Start with one narrow goal and decide whether a fixed workflow is enough. If dynamic decisions are needed, connect a model to a small set of tools, return observations to the model, impose permissions and stop conditions, and evaluate real tasks before granting more autonomy.

Make your AI workflows visible

Tornic connects your existing AI coding CLIs into visual, repeatable workflows. Keep the steps and handoffs inspectable as you build.

Explore Tornic

Related free tools