watsonx Orchestrate Agent Builder
watsonx Orchestrate Agent Builder is where an AI agent is actually defined: its name, its written instructions, the large language model that reasons for it, and the list of tools it is allowed to call. For an IBM i shop the tool list is the whole story, because it accepts a Model Context Protocol (MCP) toolkit, and that is the practical way to point an Orchestrate agent at Db2 for i data when IBM ships no IBM i connector.
What is the Agent Builder?
watsonx Orchestrate is IBM's platform for AI agents. The Agent Builder is the room inside it where you say what one agent is. IBM's developer documentation describes it as a framework for creating and importing native agents into watsonx Orchestrate using YAML, JSON or Python files. A native agent is simply one that lives in Orchestrate rather than somewhere else.
Think of hiring. You write the job title, the standing orders, and the list of systems the new hire is allowed to log into. You do not write out every task. The Agent Builder is that paperwork, and the large language model is the new hire doing the thinking.
Two IBM names sit close together here. Agent Builder is where you define an agent. The watsonx Orchestrate AI Builder is a different thing: the ADK release notes for version 2.1.0, dated 2025-12-11, record that the watsonx Orchestrate copilot was renamed to watsonx Orchestrate AI Builder.
What do you fill in on an agent?
The documentation lists the parts of a native agent. Most of them are plain text. The two that change behaviour most are the model and the tool list.
- Name, a unique identifier for the agent in the system
- Description, a human-readable summary, also used when the agent is called in as a collaborator
- LLM, the large language model that powers the agent's reasoning
- Agent style, which sets the prompting structure and the reasoning approach
- Instructions, natural language guidance such as a persona or a rule about when to use a tool
- Tools, external functions including OpenAPI definitions, Python functions and agentic workflows
- Collaborators, other agents this one can hand work to
- Knowledge base, uploaded files or a connected vector data store
There is no default model. The documentation does not name one, so you have to choose it. Agent styles have also been pruned: the ADK release notes for 2.13.0, dated 2026-07-22, say the agent styles default, react and planner are now deprecated in watsonx Orchestrate, and the current build documentation names react_core as the style that supports skills and the advanced reasoning features.
Can an agent reach Db2 for i?
Not on its own. Orchestrate has no IBM i connector and no Db2 for i connector, and nothing on IBM's Agent Builder product page names IBM i, Db2 for i or RPG. What that page does name is the pair of ways to reach data at all: OpenAPI and Model Context Protocol. MCP is the open standard for handing an AI agent a set of defined tools. So the work is to stand up an MCP server in front of your own box and add it to the agent.
There is already an MCP server for the platform. IBM publishes the IBM i MCP Server as open source under the Apache License 2.0, its tools are SQL statements written in YAML, and it talks to Db2 for i through Mapepire. Its README now carries a Stable status badge.
Wiring the IBM i MCP Server into an Orchestrate agent
- Start the server over HTTPThe IBM i MCP Server can be launched with an http transport. Its own quick start checks health at localhost port 3010 and posts to the /mcp endpoint on that port, and its Python example connects with a streamable HTTP client to that same address.
- Add it as a remote MCP toolkitOrchestrate's remote MCP command is orchestrate toolkits add with a kind of mcp, a name, a description, the server URL, a transport of either sse or streamable_http, and a tools list. A comma-separated list imports named tools and an asterisk imports all of them.
- Attach the tools to the agentBack in the Agent Builder, the imported tools become entries in the agent's tool list. From then on the agent can ask your IBM i a question it has been given a tool for, and nothing else.
Orchestrate insists you pick the transport yourself. The documentation says the ADK does not support fallback or automatic switching between protocols. It also allows 30 seconds for the server to answer during import, and the app-id flag for a stored connection applies to key-value connections only.
IBM does not document this pairing. The Orchestrate side documents a generic remote MCP toolkit and the IBM i side documents a generic MCP server; joining them is your shop's work, and the SQL in those YAML tool files is what decides how much of your business data an agent can see. Treat the tool list as a permission list.
When did the MCP pieces land?
MCP and agent changes in the watsonx Orchestrate ADK release notes
- 2025-11-25Version 2.0.0. A breaking change renames orchestrate toolkits import to orchestrate toolkits add, and MCP toolkit export arrives. Developer Edition stops requiring Docker.
- 2026-02-27Version 2.5.0 adds connection types for local MCP servers, so a toolkit can carry stored credentials.
- 2026-03-13Version 2.6.0 adds contextual metadata for MCP toolkits.
- 2026-07-22Version 2.13.0 introduces agent skills and deprecates the default, react and planner agent styles.
- 2026-09-29Version 2.18.0 adds tool shortlisting, to cut how many tools are passed to the model during planning.
Dates as printed on the ADK release notes page. Tool shortlisting matters once a toolkit has imported a long list of SQL tools from an IBM i box, because every tool offered to the model costs tokens on every turn.
No code, or the ADK?
The two ways IBM offers to build the same agent
| No-code builder | Agent Development Kit | |
|---|---|---|
| How you work | Guided flows, templates and reusable tools | Python modules and command line utilities |
| What you hand it | Answers in a form, or an uploaded standard operating procedure document | YAML, JSON or Python files |
| Adding an MCP server | Through the product interface | orchestrate toolkits add |
| Runs locally | No | Yes, through Developer Edition |
IBM's claim for the no-code route is that you can upload a standard operating procedure document and have it auto-generate a full agent workflow, logic and tools included. That is product wording, not a measured result. The ADK documentation describes itself as CLI utilities and Python modules to create, test and deploy agents and tools.
For an IBM i shop the choice is less important than it looks. Either way the agent ends up holding the same tool list, and either way somebody has to decide which SQL statements belong in it.
Related Terms
watsonx Orchestrate
IBM's AI agent and automation platform for orchestrating multi-step business workflows across enterprise applications. Since August 2026 it can launch IBM Bob from its Create an agent dialog through a Build with Bob preview, but it still ships no IBM i or Db2 for i connector.
IBM i MCP Server
An open-source, IBM-published server implementing the Model Context Protocol (MCP), letting AI agents and LLMs query Db2 for i data and take defined actions on IBM i using natural language. In beta since October 2025.
Model Context Protocol
An open standard, released by Anthropic on November 25, 2024, that defines a client-server architecture for connecting AI applications to external data sources, tools, and prompts over JSON-RPC 2.0. IBM's IBM i MCP Server is an implementation of it for Db2 for i and other IBM i resources.
Build with Bob
An option added to watsonx Orchestrate's Create an agent dialog in the August 2026 release, listed by IBM as a preview, that launches IBM Bob into a workspace wired to your Orchestrate tenant to generate agents, tools and workflows from natural language.
Db2 for i
The relational database management system built into every IBM i operating system, tightly integrated with IBM i rather than a separately installed database product.
Large Language Model (LLM)
A machine learning model trained on large volumes of text, used as the underlying technology behind generative AI tools including IBM Bob and general-purpose assistants like ChatGPT and Claude.