fix: follow moved agent evaluation guide (#2029)

Co-authored-by: Grace Lee Rui Yue <graceleeruiyue@visitor-10-59-118-36.wifi.berkeley.edu>

Author: yuelgrace1810-ops

benchmarks/datasets/mathematical-benchmarks-v1/README.md

@@ -57,7 +57,7 @@ All Jacobian treatments reach the local sidecar through
 to both jobs; only the treatment adds the Jacobian sidecar and MCP config. This
 paired setup is for workflow comparison; the public dataset is not held-out
 evidence. For a fresh task-image pull behind a proxy, follow the opt-in Buildx
-builder setup in [Run agent observations](../../../docs/how-to/run-agent-evaluations.md).
+builder setup in [Run agent observations](../../docs/run-agent-evaluations.md).
 
 To evaluate the canonical `math.find` and `math.run` surface, keep each model,
 task set, and prompt condition in a separate result root:

tests/tooling/test_harbor_egress.py

@@ -101,7 +101,7 @@ def test_agent_eval_forwards_web_search_setting_to_harbor() -> None:
 
 
 def test_agent_eval_docs_exclude_host_codex_from_the_control_protocol() -> None:
-    guide = (ROOT / "docs" / "how-to" / "run-agent-evaluations.md").read_text(
+    guide = (ROOT / "benchmarks" / "docs" / "run-agent-evaluations.md").read_text(
         encoding="utf-8"
     )