Types of AI Agents

The main types of AI agents are reflex, model-based reflex, goal-based, utility-based, and learning agents. Most developers now add a language model that picks tools at runtime. The practical choice is therefore which decision loop the agent runs. The type depends on how the agent picks its next step.

The classic agent types

Reflex agents choose an action from the current input alone. A thermostat that switches heating on below a set temperature is a reflex agent. It keeps no memory of earlier states.

Model-based reflex agents keep an internal state. That state tracks facts the agent cannot observe directly, such as a previous reading or an open session. The agent combines this state with its rules before it acts.

Goal-based agents pick actions that move them toward a target state. They search or plan over possible actions to reach the goal. A route planner that finds a path to a destination works this way.

Utility-based agents score possible outcomes and pick the action with the highest expected score. They handle trade-offs, such as cost against speed or accuracy against latency. A pricing engine that weighs margin against conversion behaves like this.

Learning agents improve from feedback. They contain a performance element that acts and a learning element that updates the rules or the model. A recommendation system that adjusts after user clicks is one example.

Agent types built on language models

Language-model agents usually combine three parts: a model, a set of tools, and a loop. The model reads the task, chooses a tool, reads the result, and decides the next step. Developers call this a tool-calling loop.

A single-tool agent calls one function per task. It suits narrow jobs, such as looking up an order status. Its failure modes are few because the decision space is small.

A ReAct-style agent alternates between short reasoning steps and tool calls until it reaches an answer. Each observation feeds the next step. This type handles open questions that need several lookups in sequence.

A planner-executor agent writes a plan first, then runs each step with a narrower worker. The planner stays out of the tool loop. This design reduces wasted calls on long tasks, although a poor plan still produces poor results.

A multi-agent system splits work across specialist agents. An orchestrator assigns subtasks and merges the results. Use this type when subtasks need separate tools or separate permissions. Otherwise, one agent with clear instructions is easier to debug.

Workflow-driven agents follow a fixed graph of steps, and the model fills in content at each node. Autonomous agents choose their own path through the task. Workflow-driven agents are easier to test. Autonomous agents cover more inputs but need more review.

Choosing a type for your build

Start with the decision the agent must make. Match the type to that decision. Three questions narrow the choice:

• Does the next step depend on earlier steps? If not, a reflex or single-tool agent fits.
• Do the steps follow a known order? If yes, a workflow-driven design fits.
• Does the agent trade cost against quality? If yes, add a utility score to its choices.

Keep the first version small. Add memory, planning, or extra agents only after a test shows that the simpler design fails on real inputs.

Giving each agent type tools through MCP

Every type above needs tools to act on the outside world. The Model Context Protocol gives an agent a standard way to discover and call tools. APIbase is an MCP gateway. It offers a large catalog of tools from many providers across many categories. An agent reaches these tools through one interface instead of one integration per provider.

Pricing uses a range. The current price range is 0 to 29.99 USD. Check the live price before you budget a build, because a tool-calling loop multiplies calls.

Tool access does not change the agent type. A reflex agent with one MCP tool stays a reflex agent. Planning, memory, and evaluation remain in your code.

Limits and failure modes

Loops cost money. A tool-calling agent can repeat calls when a result stays unclear. Set a step limit and a spending limit before the agent runs in production.

Permissions need the same care. Give each agent only the tools its task requires. A research agent should not hold write access to billing.

Test each type on real inputs. Log every tool call and its result. Read those logs when the agent returns a wrong answer, because the failing step usually appears there first.

Live pricing — developer

Connect via MCP

$ curl -X POST https://apibase.pro/api/v1/tools/account.timeseries/call \
  -H "Content-Type: application/json" -d '{"params": {}}'

FAQ

What are the main types of AI agents?

The classic types are reflex, model-based reflex, goal-based, utility-based, and learning agents. Language-model agents add practical types such as single-tool agents, ReAct-style agents, planner-executor agents, and multi-agent systems.

Which type should a developer build first?

Build the simplest type that fits the decision. A single-tool agent or a workflow-driven design often works for a first version. Add planning or extra agents only after tests show that the simpler design fails.

Does adding MCP tools change the agent type?

No. Tools give an agent more actions, but the type depends on how it chooses its next step. Planning, memory, and evaluation remain in your application code.

How do I control the cost of a tool-calling agent?

Set a maximum number of steps and a spending limit for each run. Log every tool call so you can see where repeated calls occur. Check the current price range before you budget, which is 0 to 29.99 USD.

Recommended next step

1. STEP 1/2 Connect your MCP client
2. STEP 2/2 Browse the full catalog

Related guides

• serper api
• desktop commander mcp
• usdc coin
• power bi mcp

Published: 2026-10-28Source: live apibase.pro catalog snapshot