Apple can fairly say: “Agents shouldn’t directly edit our serialization format.” But if they advertise agentic workflows and the agent can’t reliably do core build/debug tasks (target build settings, linking, schemes), then the product gap is real. Agentic workflows require a closed loop: apply build-setting changes → build → fix → repeat. If the only exposed interface is “chat + inline code edits,” it’s not agentic for build/debugging. Therefore Apple needs a supported automation surface (typed project actions with preview/undo), regardless of how ugly project.pbxproj is. Expected: Agent can apply target/scheme/build-setting changes (typed + preview/undo), run build, iterate on failures. Actual (Xcode 26.3 RC): Agent mostly suggests or edits source files; can’t reliably adjust the build/ debug configuration loop for multi-target/C++ integration. C/C++ integration is where an agent should shine because it’s repetitive, error-driven, and configuration-heavy. A “real” agentic C++ loop in Xcode would need to be able to do things like (for a specified target + config): Set HEADER_SEARCH_PATHS, USER_HEADER_SEARCH_PATHS, LIBRARY_SEARCH_PATHS, FRAMEWORK_SEARCH_PATHS Add OTHER_CPLUSPLUSFLAGS, OTHER_LDFLAGS, GCC_PREPROCESSOR_DEFINITIONS Choose CLANG_CXX_LANGUAGE_STANDARD / CLANG_CXX_LIBRARY Add/link .a/.dylib/.framework, set “Link Binary With Libraries” Add “Run Script” phases (e.g., copy deps, run CMake, generate headers) Build a named scheme/destination and iterate until xcodebuild is green If Xcode’s agent can’t apply those (only suggest UI clicks), it’s basically missing the highest- leverage use case. Apple’s “agentic” vision needs a first-party, schema/intent layer (or an official project-model editing API with diff/undo) to make build/debug automation a first-class workflow—not just code completion. If Xcode integration can’t reliably handle project/build edits in one pass, it loses the main advantage over terminal workflows.

If an “agent” can’t: run a build/test for a specified scheme/target/destination, interpret failures, apply the needed project/build-setting changes, iterate until green, then it’s an in-editor assistant, not an agentic workflow.