diff --git a/.cargo/config.toml b/.cargo/config.toml
index 1d9c9ecc9da..26a4fe335ac 100644
--- a/.cargo/config.toml
+++ b/.cargo/config.toml
@@ -9,6 +9,12 @@ debug = true
 codegen-units = 16
 lto = "thin"
 
+[profile.release-no-lto]
+inherits = "release"
+lto = false
+# Prioritize compile time when LTO is not relevant to the measurement.
+codegen-units = 16
+
 [profile.bench]
 inherits = "release"
 lto = "thin"
diff --git a/AGENTS.md b/AGENTS.md
index 2003d6dba10..cdca99ffae8 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -41,6 +41,10 @@ Key technical traits: async-first (tokio), Arrow-native, versioned writes with m
 * Coverage: `cargo +nightly llvm-cov -q -p <crate> --branch`
 * Coverage HTML: `cargo +nightly llvm-cov -q -p <crate> --branch --html`
 * Coverage for file: `python ci/coverage.py -p <crate> -f <file_path>`
+* Use repository-defined Cargo profiles instead of ad hoc LTO overrides.
+* Use `release` or `bench` for CPU-bound benchmarks and performance numbers intended for comparison.
+* Use `release-with-debug` for optimized profiling that needs debug symbols.
+* Use `release-no-lto` only for local debugging, IO-bound benchmarks, or compile-time-sensitive performance investigation where LTO would not affect the measured bottleneck.
 
 ### Python / Java