diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 0000000000..36d0d7841d
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,49 @@
+# DXC agent guidance
+
+## Investigating compile-time regressions and miscompiles
+
+- Start from evidence. Identify the exact pass, hot function, and triggering
+  operation before proposing a fix. Enumerate every limit, cutoff, cache, and
+  heuristic on that path.
+- Check this fork before assuming an LLVM design flaw. DXC is based on LLVM
+  3.7, and local divergences are commonly, but not always, marked
+  `HLSL Change`. Reconstruct the local pre-change behavior before reaching
+  outside the repository.
+- Inspect local history before designing a new mechanism. Run `git blame` on
+  the hot line and use `git log -p -- <file>` and
+  `git log -S<symbol> -- <file>` to find when and why it changed. Treat a
+  surprising tuning constant as a regression candidate first. Use
+  `git show <commit>^:<path>` to inspect the pre-change file. If a pristine
+  upstream comparison is still needed, consult the `llvmorg-3.7.1` tag in
+  `llvm/llvm-project`.
+- Check the scope of shared analysis settings. A file-scope limit changed for
+  one pass can affect every consumer of that analysis. Enumerate the users and
+  state the affected passes in the PR.
+- Prefer restoring an upstream default over adding a DXC-specific limit or
+  heuristic to core LLVM code. Require stronger evidence for a new mechanism
+  because this fork has limited regression coverage for LLVM analyses and
+  transforms.
+- Account for the test gap. `test/Transforms/lit.local.cfg` and
+  `test/Analysis/lit.local.cfg` disable discovery with
+  `config.suffixes = []`, including the ordinary GVN and
+  MemoryDependenceAnalysis suites. Add focused coverage under
+  `tools/clang/test/DXC/Passes/`, but do not mistake a targeted test for broad
+  pass-regression coverage.
+- Build reproducers from the original failure's structural characteristics,
+  not from the proposed fix. Verify that the repro has the same sensitivity to
+  relevant limits and options before trusting it.
+- Treat surprising experimental results as new evidence. If an unrelated knob
+  outperforms the proposed fix, stop and re-derive the root cause instead of
+  continuing the existing plan.
+- Use cross-fork comparisons for "how does upstream work now?" and local
+  history for "what did DXC change?" Do the cheaper local-history check first.
+
+### Example
+
+`MemoryDependenceAnalysis` used a file-scope `BlockScanLimit` of 500 after
+commit `08f3100f2503b8d6750c1b65b31fe6ae055be84f` / #2725 raised LLVM 3.7's
+default of 100 for a DSE case. The accompanying `-memdep-block-scan-limit`
+option was routed to DSE, but the shared default also affected GVN's
+memory-dependence queries. #9003 restores the upstream default and adds a
+focused GVN test after the divergence caused a severe compile-time regression
+on a large generated compute shader.