An AI agent builder sets up a model with instructions and tools

An AI agent builder is the place where you give an AI model its instructions and its tools. OpenAI’s Agents SDK docs define an agent as a model with instructions and tools. Oracle’s NetSuite docs split the idea in two. A skill does one kind of job well. An agent is a decision maker that uses skills to reach a goal.

Builders come in two shapes. Some are canvases you click through. Business system vendors such as Salesforce, Microsoft, SAP and HubSpot sell them, and they work on data your team already keeps there. Others are code libraries and cloud platforms. We list both below, with the facts each vendor publishes, and then say when a custom agent fits better. We left out general no-code tools that are not tied to a business system. Joule Studio’s n8n workflows are covered below.

Builders from business system vendors work on the data already there

These builders work on data that already sits in one vendor’s system. They fit best when the whole job happens there.

Salesforce Agentforce Builder

Agentforce Builder connects agents to Flows, Apex, MuleSoft APIs and prompt templates. Agent Script is its new language for building agents, and it lets you set conditions, loops and transitions. Prebuilt blocks cover jobs such as order management and general FAQs. With A2A support, Agentforce can securely connect to third-party agents. You can find trusted MCP servers in AgentExchange and deploy the one you pick with no code. Agentforce Developer Edition is a free space to prototype. Our team has built Salesforce agents in Agentforce and as custom agents through Salesforce’s API (Salesforce AI agents).

Microsoft Copilot Studio

Microsoft calls Copilot Studio a graphical, low-code studio for building and managing agents and workflows. You can create an agent by describing it in plain language, then test it before you publish. Some agents can get their own account and work on tasks without being asked. Its workflows include built-in testing and steps where a person signs off. Copilot Studio reaches other data through prebuilt or custom connectors, and it supports MCP tools and resources. Agents use Copilot Credits, and unused credits do not carry over to the next month.

SAP Joule Studio

SAP describes Joule Studio as a tool for building and running AI agents and workflows. It can build multistep n8n workflows that direct agents, and custom agents can act across SAP and non-SAP systems. You need Joule deployed. You also need an SAP BTP Enterprise Agreement (BTPEA), a Cloud Platform Enterprise Agreement (CPEA) or pay-as-you-go SAP BTP. Design time is free. A time-bound offer lets custom agents run at no extra cost through the end of 2026, about six weeks after we read it. After that, SAP bills production agents on compute, storage, agent memory, workflow runs and AI tokens. Each Joule agent gets its own identity, and it can do only what the end user may do.

HubSpot agent builder

Custom agents use HubSpot Credits every time they run. Workflows use none. HubSpot also says a custom agent’s actions vary from one run to the next. The agent builder was called Breeze studio before, and it comes with the Starter, Professional and Enterprise editions. It builds custom agents and workflows on one canvas with no coding. Our team has built HubSpot AI agents, including agents built on Anthropic’s Claude (HubSpot AI agents).

Odoo 20 AI agents

Odoo’s docs warn that an agent started by an automated action creates or updates records without asking for approval. An Odoo agent has four parts: instructions, knowledge, skills and tools, and triggers that decide when it runs. An agent with no skills can only give information. It cannot change the database. An agent can call other agents as subagents. Installing the AI app on Odoo’s Standard plan moves the database to the Custom plan, and AI use spends IAP credits.

Frappe Flow, for ERPNext

A Flow trigger can run an agent when a document changes or on a schedule, and then the agent makes every change without asking. Flow puts an AI assistant and an agent framework into a Frappe site, from Frappe v16.60. It is in beta. Frappe says it can make mistakes. Only a System Manager can create agents. The built-in Flow agent stops at 500 steps, a cap you cannot change. Flow works with any model server that speaks the OpenAI API, including Ollama and LM Studio.

Oracle NetSuite: SuiteAgents, and tools for your own agent

NetSuite Next adds Oracle’s own agents, and some agent capabilities are only available there. They run through Ask Oracle and through specific NetSuite pages or workflows. They cannot reach records, reports, pages or actions your role cannot reach. System Notes logs changes made by SuiteAgents under the agent’s name. The Oracle pages we read do not say whether customers can build their own SuiteAgent. Your own agent reaches NetSuite from outside, through the AI Connector Service. Oracle’s MCP Standard Tools SuiteApp is one route. Custom MCP tools are the other. Each needs a SuiteScript 2.1 script, a JSON schema and a toolset SDF object. A tool script cannot use the N/http, N/https, N/llm or N/sftp modules. For NetSuite’s built-in AI, you cannot bring your own model.

Developer kits and cloud platforms work outside any one business system

These kits reach a system through code and its APIs, and most of them need a developer. They fit when a job crosses systems, or when you want control of each step.

OpenAI Agents SDK

OpenAI calls the SDK a production-ready upgrade of Swarm, its earlier agent test. Its building blocks are agents, handoffs between agents, and guardrails that check inputs and outputs. It has built-in steps for bringing a person into a run. It can give agents remote MCP tools alongside your own functions. OpenAI says the SDK leaves deployment, storage, approvals and runtime in your app’s hands.

OpenAI Agent Builder, scheduled to close November 30, 2026

Agent Builder is a visual canvas for multi-step agent workflows. It shows up in search results for this term, but OpenAI is deprecating it, and it is scheduled to shut down on November 30, 2026. ChatKit stays available. A workflow can also be downloaded as SDK code and run on your own servers, which keeps a live workflow going after the shutdown date.

Anthropic Claude Agent SDK

Anthropic’s Agent SDK gives you the same agent loop and tools that power Claude Code. It runs in Python and TypeScript. Anthropic also offers Managed Agents, which runs the loop for you, in a sandbox Anthropic hosts or one on your own servers.

Google ADK and Gemini Enterprise Agent Platform

Google’s ADK is an open-source framework for building agents, in Python, TypeScript, Go, Java and Kotlin. It has adapters for many other models, including models that run on your own machines. On Gemini Enterprise Agent Platform, Agent Runtime deploys and scales agents for you, and keeps chat state in Sessions and Memory Bank.

Microsoft Foundry Agent Service

Foundry Agent Service is Microsoft’s managed platform for deploying and scaling agents. It offers three agent types. Prompt agents run with no code or infrastructure to manage. Voice-based prompt agents handle real-time voice, and hosted agents run your own code. Hosted agents can be built with Agent Framework, LangGraph, the OpenAI Agents SDK, the Anthropic Agent SDK or your own code. Agents reach MCP servers through the MCP tool. A private MCP server needs private networking and its own subnet.

Amazon Bedrock AgentCore

AgentCore is Amazon’s platform for building and running agents with any framework and model. Its Gateway turns APIs and Lambda functions into MCP tools. Its Policy service works with Gateway. It checks every tool call that passes through Gateway before the call runs. AgentCore also runs a managed agent loop. Each of its sessions runs in its own isolated microVM, a small virtual machine. Sign-in works with Cognito, Okta, Microsoft Entra ID and Auth0. Amazon has closed new sign-ups for its older Bedrock Agents.

How the builders differ on approval and reach

The table compares what each vendor says about a person approving changes, and the route each builder uses to reach other systems. Where a vendor’s page says nothing on a point, the cell says so.

Scroll the table sideways to see every column.

Builder What the vendor says about approval How it reaches other systems
Agentforce Builder Guardrails mix your own safeguards with Salesforce’s. An Audit Trail tracks agent actions Flows, MuleSoft APIs, A2A and MCP servers
Copilot Studio Workflows have human-in-the-loop controls Prebuilt or custom connectors, and MCP tools and resources
Joule Studio Agents can do only what the end user may do SAP and non-SAP systems, and n8n workflows
HubSpot agent builder Actions vary across runs. In the agent tools beta, an agent has no CRM data unless a tool gets it Agent tools, which are custom workflow actions, in beta
Odoo AI agents Agents started by automation act without approval Outside agents can connect through Odoo’s MCP server
Frappe Flow Triggers make every change without asking Not stated on the page we read
OpenAI Agents SDK Built-in human-in-the-loop steps Your functions and remote MCP tools
Google ADK Not stated on the page we read Not stated on the page we read
Foundry Agent Service Not stated on the page we read The MCP tool, to public or private servers
Bedrock AgentCore Policy checks each tool call through Gateway before it runs Gateway turns APIs into MCP tools. Runtime supports MCP and A2A

Two rows say more about the model than about reach. Frappe Flow takes any model server that speaks the OpenAI API, and Google’s ADK has adapters for many models and providers.

When a custom agent beats a builder

A custom agent wins when the job crosses systems, when every write needs a person, or when the builder will not take your model or rules. Here are the five signs we look for, each tied to a vendor fact above.

  1. The job spans two systems. In HubSpot, an agent tool is a custom workflow action made available to agents, and agent tools are in beta. A quote that needs stock from the ERP and a price rule from the CRM needs either such a tool or an agent that can call both systems.
  2. Every write needs a person. Frappe Flow triggers and Odoo’s automated agents change records without asking. A custom agent can save only drafts, so a person approves each one. In the multi-agent RFQ system our team built in ERPNext and Frappe CRM, separate agents handle intake, item matching, pricing, supplier RFQs and reading supplier replies, and a person approves every quote (case study).
  3. The builder cannot reach out. A NetSuite custom tool script cannot call an outside service. NetSuite’s built-in AI will not take your own model.
  4. You need the same steps every time. HubSpot says custom agent actions vary across runs, while its workflows do not use credits. Fixed steps belong in a workflow or in code, with the AI used only where reading or judging is needed.
  5. The builder is closing or changing. OpenAI’s Agent Builder is scheduled to shut down on November 30, 2026, twelve days after we read the page. Amazon has closed new sign-ups for Bedrock Agents. Keep your agent’s rules in a form you can move.

Our comparison of Agentforce, Joule, Breeze and a custom agent goes deeper on the three CRM and ERP vendors. For the build itself, see how to build an AI agent for an ERP and our AI agent development service.

Sources

Every source behind a figure or a vendor fact in this article, with the day we read it.