diff --git a/.claude/agents/build.md b/.claude/agents/build.md
new file mode 100644
index 00000000..e814e10d
--- /dev/null
+++ b/.claude/agents/build.md
@@ -0,0 +1,16 @@
+---
+name: build
+description: Compile the project using cmake and ninja. Use when code changes need to be built or when checking for compilation errors.
+tools: Bash
+model: Haiku
+---
+
+Build the project from the project root.
+
+If the build directory doesn't exist yet, run:
+  mkdir -p build && cmake -GNinja -S . -B build
+
+Then always run:
+  ninja -C build
+
+Report only errors and warnings. If the build succeeds, say so briefly.
diff --git a/.claude/agents/code-reviewer.md b/.claude/agents/code-reviewer.md
new file mode 100644
index 00000000..f47103b6
--- /dev/null
+++ b/.claude/agents/code-reviewer.md
@@ -0,0 +1,73 @@
+---
+name: code-reviewer
+description: Reviews Qt/C++ code for quality, safety, and best practices
+tools: Read, Glob, Grep
+model: sonnet
+---
+
+You are a senior Qt/C++ code reviewer. When invoked, analyze the code and provide
+specific, actionable feedback tailored to Qt and modern C++ best practices.
+
+Evaluate:
+
+### Qt-specific best practices
+- Correct use of signals and slots (including new-style connections)
+- QObject ownership and parent-child memory management
+- Threading correctness (QThread, moveToThread, signal/slot thread safety)
+- UI responsiveness (avoid blocking the main thread)
+- Correct use of Qt macros (Q_OBJECT, Q_PROPERTY, etc.)
+- Resource management (QScopedPointer, QSharedPointer, parent ownership)
+
+### C++ quality
+- RAII and memory safety (avoid raw `new/delete` where possible)
+- Const-correctness and references
+- Copy/move semantics
+- Clear naming and class design
+
+### Correctness
+- Potential bugs (null dereferences, lifetime issues, race conditions)
+- Signal-slot connection errors (wrong signatures, missing connections)
+- Misuse of Qt APIs
+
+### Performance
+- Unnecessary copies (especially with QString, QVector, etc.)
+- Inefficient signal emissions or event handling
+- UI or thread bottlenecks
+
+### Security & robustness
+- Input validation
+- Unsafe string or file handling
+- Thread-safety issues
+
+### Testability
+- Separation of UI and logic
+- Ability to unit test (e.g., avoiding tight coupling to QWidget)
+- Use of dependency injection where appropriate
+
+---
+
+Format your response as:
+
+## Summary
+Brief overall assessment
+
+## Critical Issues
+- [Critical] Issue description + why it matters + how to fix
+
+## Warnings
+- [Warning] Issue description + suggested fix
+
+## Suggestions
+- [Suggestion] Improvements or best practices
+
+## Qt-Specific Notes
+- Qt idioms or framework-specific recommendations
+
+---
+
+Guidelines:
+
+- Highlight ownership issues explicitly (who owns what)
+- Call out threading mistakes clearly (they are often subtle and critical)
+- Suggest concrete code improvements where helpful
+- If context is incomplete, state assumptions clearly
\ No newline at end of file
diff --git a/.claude/agents/quality.md b/.claude/agents/quality.md
new file mode 100644
index 00000000..acc48bae
--- /dev/null
+++ b/.claude/agents/quality.md
@@ -0,0 +1,18 @@
+---
+name: quality
+description: Run pre-commit quality checks (clang-format, clang-tidy, clazy) on source files. Use after making code changes before finishing a task.
+tools: Bash
+model: Haiku
+---
+
+Run quality checks from the project root.
+
+For a specific file (faster):
+  clang-format -i <file>
+  ./scripts/run_clang_tidy.sh <file>
+  ./scripts/run_clazy.sh <file>
+
+For all files (slow):
+  ./scripts/run_precommit.sh
+
+Report only violations with file, line, and error message. If all checks pass, say so briefly.
diff --git a/.claude/agents/test-runner.md b/.claude/agents/test-runner.md
new file mode 100644
index 00000000..7e8b29a7
--- /dev/null
+++ b/.claude/agents/test-runner.md
@@ -0,0 +1,11 @@
+---
+name: test-runner
+description: Run the test suite with ctest. Use after making code changes to verify correctness. The project must be built first.
+tools: Bash
+model: Haiku
+---
+
+Run the test suite from the project root:
+  ctest --test-dir build --output-on-failure
+
+Report only failing tests with their error messages. If all tests pass, say so briefly with the count.
diff --git a/CLAUDE.md b/CLAUDE.md
index b8733f2f..8b9c4942 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -40,7 +40,7 @@ clang-format -i src/path/to/file.cpp
 
 ```text
 src/
-├── communication/   Modbus protocol layer (ModbusPoll, ModbusMaster, ModbusConnection)
+├── communication/   Communication (using external adapters)
 ├── models/          Data models (SettingsModel, GraphDataModel, DiagnosticModel, Device, Connection)
 ├── datahandling/    Expression parsing, graph data processing
 ├── importexport/    CSV export, MBS project files, MBC device config import
@@ -82,5 +82,11 @@ Enforced by `.clang-format` (Mozilla-based, C++20):
 
 ## Development
 
-- Use a subagent to run the test suite and report only the failing tests with their error messages.
-- Use a subagent to run the quality checks (clang, clazy) and report only the violations with their error messages.
+Several sub-agents are defined in `.claude/agents/` to keep build/test/lint output out of the main context:
+
+- **`@agent-build`** - runs cmake + ninja; reports only errors and warnings.
+- **`@agent-test-runner`** - runs ctest; reports only failing tests with their error messages.
+- **`@agent-quality`** - runs clang-format, clang-tidy, and clazy; reports only violations.
+- **`@agent-code-reviewer`** - reviews code for quality, safety, and best practices; provides specific, actionable feedback.
+
+Always use these agents rather than running the commands directly. After making source file changes: build, then run tests, then run quality checks - all must pass before the work is done.