What coding agents do

A code assistant suggests; a coding agent acts. That's the difference that matters for quality and sign-off. Predictive code completion offers the next line as ghost text, and you decide. An agent takes a goal such as "add a validation to the core agreement item and prove it works", breaks it into steps, calls tools, reads the results and keeps going until it thinks it's done. It can create objects, edit source, activate and run tests without a person approving each keystroke.

The agentic loop

SAP's own help page, "Scenario: Agentic Loop", walks through a coding agent creating an ABAP class through the ADT MCP server. The agent lists the available systems, finds the right object type, validates the name and package, creates a transport request, creates the class, writes the implementation, activates it and runs ABAP Unit. SAP adds a note that the number and order of tool calls may vary with the MCP host. That variation is the point: the agent decides the sequence, not you.

What MCP adds

The Model Context Protocol (MCP) is an open standard for connecting AI applications to tools and data. An MCP host, such as an agent in VS Code or Claude Code in a terminal, connects to MCP servers. Each server publishes a list of tools with names and input schemas, and the agent decides which tool to call. Before MCP, every agent needed a custom integration for every system. With MCP, SAP builds one server and any MCP-capable agent can use it. SAP describes the ADT MCP server in exactly those terms: an open channel that lets the coding agent of your choice work with ABAP code and systems.

For ABAP this matters more than for most languages. ABAP source lives on the server, not on your disk, and a change isn't real until it's activated in the system. A generic agent that only edits local files can't activate a class, run ABAP Unit or see an ATC finding. The ADT MCP server gives the agent those system-side actions, using the developer's own logon, so the agent works with the same authorizations as the person.

Context: telling the agent your rules

An agent starts every session knowing nothing about your system landscape, naming rules or transport habits. SAP's documentation recommends an AGENTS.md file, which it calls an industry-standard configuration file, and provides a template: use cloud-compliant ABAP syntax and APIs, use a named package and prefix, use the ADT MCP server to create objects, always run unit tests after a change, put tests in the test class include. Claude Code reads a CLAUDE.md file for the same purpose, and can also read AGENTS.md. Course sap-aid-203 builds one for a real ABAP repository.

Both files are guidance, not enforcement. Anthropic's Claude Code documentation says Claude treats instruction files as context, not enforced configuration. The hard limits come from the tools you expose, the authorizations of the user the agent works as, permission rules, automated checks such as ATC, and human review.

Approving tool calls

Most MCP hosts let you decide which tools an agent may call and when it must ask. SAP's "Using MCP Tools" page describes the VS Code flow: after the connection is established you see the ADT MCP tools, select or deselect them for the current request, switch to agent mode and type the prompt. When a tool needs approval, a dialog shows the tool details, and you can approve it once, for the session, for the workspace or for all future calls. Those choices are the first line of control. A sensible starting point is to approve read-only tools such as search, read and explain freely, and to approve create, edit, activate and transport tools one call at a time until the team trusts the setup.