AI Subagents for Optimized ABAP Development¶

AI Subagents are specialized AI assistants, each focused on one type of ABAP task (finding objects, reading code, running analysis, etc.). Instead of one general AI doing everything, subagents split work across focused specialists.

Why this matters:

Better results — a dedicated code reviewer catches more issues than a general assistant juggling multiple goals
Longer conversations — heavy operations run in separate context windows, so your main chat stays responsive
Lower cost — simple tasks (search, read) use cheaper/faster models; complex tasks use smarter ones

Available Subagents¶

Agent	What it does	Tier
`abap-orchestrator`	Routes tasks, writes all code, coordinates other agents	3 (Premium)
`abap-code-reviewer`	Deep code review — security, performance, best practices	3 (Premium)
`abap-usage-analyzer`	Where-used analysis, dependencies, change impact	2 (Mid-tier)
`abap-quality-checker`	ATC analysis, unit tests, code health	2 (Mid-tier)
`abap-historian`	Version history, transport requests	2 (Mid-tier)
`abap-debugger`	Runtime debugging — breakpoints, stepping	2 (Mid-tier)
`abap-troubleshooter`	Analyze dumps, traces, performance issues	2 (Mid-tier)
`abap-data-analyst`	Query SAP tables, analyze data patterns	2 (Mid-tier)
`abap-discoverer`	Find ABAP objects by name/pattern	1 (Cheap/Fast)
`abap-reader`	Read and extract info from source code	1 (Cheap/Fast)
`abap-creator`	Create new ABAP objects (shells)	1 (Cheap/Fast)
`abap-visualizer`	Create diagrams from code	1 (Cheap/Fast)
`abap-documenter`	Generate technical documentation	1 (Cheap/Fast)

How to Use Subagents¶

In GitHub Copilot Chat, type @abap-orchestrator to start. The orchestrator is the only agent exposed directly in the chat dropdown — it calls other agents automatically as needed.

@abap-orchestrator analyze ZCL_ARTICLE_HANDLER and suggest improvements

For example, the orchestrator might:

Delegate "find related classes" → abap-discoverer (cheap, fast)
Delegate "read the code" → abap-reader (cheap, fast)
Delegate "usage analysis" → abap-usage-analyzer (mid-tier)
Synthesize findings and write recommendations itself (premium)

You can also invoke other subagents directly with @agent-name if needed. Ask Copilot to make an agent available in the dropdown — it can update the agent's .agent.md file to enable this.

Setup¶

Subagent configuration is stored at the workspace level in .vscode/settings.json and .github/agents/. Each project can have its own configuration.

In normal usage, you do not need to edit these files manually. Copilot can configure models, generate/update agent files, validate them, and enable/disable subagents through chat commands.

Step 1 — Configure models¶

Ask Copilot:

Configure subagents for ABAP development

Copilot will suggest models for each tier and ask for confirmation before applying. Recommended assignments:

Tier	Agents	Example models
1 — Cheap/Fast	discoverer, reader, creator, visualizer, documenter	Claude Haiku 4.5, Gemini 3 Flash
2 — Mid-tier	usage-analyzer, quality-checker, historian, debugger, troubleshooter, data-analyst	GPT-4o, Claude Sonnet 4
3 — Premium	orchestrator, code-reviewer	Claude Sonnet/Opus 4.6, GPT-5.4

Avoid assigning premium models to Tier 1 agents — it eliminates the cost benefit without improving results for simple tasks.

Step 2 — Enable subagents¶

Ask Copilot:

Enable subagents

This creates agent files in .github/agents/ and validates them.

Step 3 — Allow agent delegation (if prompted)¶

You may see a notification asking to enable chat.customAgentInSubagent.enabled. Click Enable Setting — this allows the orchestrator to call other agents.

Managing Subagents¶

All management is done through Copilot chat:

What you want	What to ask
Check current status	`Show subagent status`
Disable all agents	`Disable subagents`
Re-enable agents	`Enable subagents`
Change a model	`Change abap-discoverer to use GPT-4o`
See available models	`What models can I use for subagents?`
See available tools	`List available tools for subagents`

When you disable subagents, agent files move to agents_disabled/ (not deleted). Re-enabling restores them with your customizations intact.

Customizing Agent Tools¶

Each agent's .agent.md file in .github/agents/ defines which tools it can use. You can edit these files directly or ask Copilot to do it:

Add the abap-trace tool to abap-troubleshooter

Changes survive disable/re-enable cycles — only the model: line is updated when you change models.

✅ User Control: You decide which models to use for each agent tier

What to Be Aware Of¶

⚠️ Model Availability: Some models shown in the list may not work (e.g., "GPT-4o mini"). The system validates and auto-disables if errors are detected.

⚠️ VS Code Setting Required: chat.customAgentInSubagent.enabled must be true for delegation to work, otherwise main agent's model may be used for all subagents which can result in a lot of premium request usage.

⚠️ Workspace-Specific: Settings and agent files are per-workspace, not global

⚠️ Agent Files in Git: The .github/agents/ folder will appear in your version control - add to .gitignore if you don't want to share

⚠️ Frequently-Used Agents: Agents like abap-discoverer and abap-reader get called often - using expensive models for these defeats the cost benefit

Troubleshooting¶

"Cannot enable subagents - missing models"¶

All 13 agents must have models configured. Ask Copilot to configure missing agents.

Agent files show validation errors¶

Some model names aren't valid for agent files. Try a different model (e.g., use Claude Haiku 4.5 instead of GPT-4o mini).

Subagents auto-disabled¶

This happens when configured models become unavailable. Reconfigure with available models.

Ghost files in explorer after disable¶

This is a VS Code refresh issue. The extension refreshes the explorer automatically, but occasionally you may need to collapse/expand the folder.

Delegation not using custom agents¶

Make sure chat.customAgentInSubagent.enabled is set to true in your VS Code settings.