diff --git a/backend/AGENTS.md b/backend/AGENTS.md
index 46f6aa0296..260ea508ef 100644
--- a/backend/AGENTS.md
+++ b/backend/AGENTS.md
@@ -37,33 +37,27 @@ Image result URL policy: [image search guide](packages/harness/deerflow/communit
 
 ### Backend Benchmarks
 
-`scripts/benchmark/context_snapshot/`: explicit `run-live` needs provider env
-vars; `summarize` and pytest are offline. See its README for the protocol.
+Keep `scripts/benchmark/` standalone and reproducible. Import production
+functions and reuse their runtime.
 
-Benchmarks in `scripts/benchmark/` must be standalone and reproducible. Import
-production functions; never duplicate them or introduce an alternative runtime.
-
-`scripts/benchmark/context_memory_comparison/` runs an offline Gateway with
-SQLite. Keep original- and fresh-thread results separate.
-
-Sandbox benchmarks reuse production lifecycle. Reject negative/NaN timing values
-before provider creation, export no secrets, and fail on final cleanup errors.
-See the sandbox benchmark README for supported options and tests.
-
-- Pin every external dataset by immutable revision and SHA-256. Callers provide
-  the local dataset path; evaluation commands must not silently download data.
-- Never commit upstream dataset text, credentials, complete provider requests,
-  or response headers. Committed manifests may contain stable IDs and source
-  locators. Synthetic cases must identify themselves as synthetic.
-- Read provider credentials and endpoints from named environment variables.
-  Version model IDs, inference parameters, prompts, retry rules, clocks, and
-  random seeds in the evaluation config.
-- Public raw results may contain case IDs, policy decisions, model hypotheses,
-  grades, and non-secret response metadata. Keep dataset questions, reference
-  answers, memory content, and full provider payloads in ignored local run
-  directories.
-- Use fixed clocks and deterministic ordering for offline selection. Results
-  must record the config, manifest, prompt, dataset, and git revisions used.
+- `context_snapshot` needs provider env vars for `run-live`. See its README.
+  `summarize` and pytest run offline.
+- `context_memory_comparison` runs Gateway offline with SQLite. Report
+  original- and fresh-thread results separately.
+- Sandbox benchmarks use the production lifecycle. Reject negative or NaN
+  timing values before provider creation, export no secrets, and fail on
+  cleanup errors. See the sandbox benchmark README for options and tests.
+- Pin datasets by immutable revision and SHA-256. Require a caller-provided
+  local path without automatic downloads. Label synthetic cases. Manifests
+  may include stable IDs and source locators.
+- Read provider credentials and endpoints from named env vars. Version model
+  IDs, inference parameters, prompts, retries, clocks, and seeds in eval config.
+- Public raw results may include case IDs, policy decisions, model hypotheses,
+  grades, and non-secret response metadata. Keep dataset text, reference
+  answers, memory, and full provider payloads in ignored local run directories.
+  Exclude credentials and response headers from commits.
+- Use fixed clocks and deterministic offline ordering. Record config, manifest,
+  prompt, dataset, and git revisions in results.
 
 `scripts/benchmark/deermem_eviction/` evaluates the production
 `select_facts_for_capacity()` implementation used by DeerMem. It compares only
@@ -84,15 +78,12 @@ The offline test suite must not require network access, provider credentials,
 or the LongMemEval dataset. Small LongMemEval-shaped fixtures must be synthetic
 and generated by tests.
 
-`scripts/benchmark/concurrency/` measures multi-process contention on the
-`users` table (N separate OS processes, not asyncio tasks) for SQLite vs
-Postgres -- the scenario `CONFIGURATION.md` requires Postgres for. `worker.py`
-connects directly via SQLAlchemy (skipping the ~8.5s Alembic bootstrap the
-orchestrator already ran once) and mirrors the app's per-connection SQLite
-PRAGMAs; `run_concurrency_bench.py` seeds a disposable per-run Postgres schema,
-synchronises workers on a READY/GO barrier before timing, and exits non-zero on
-any crash, short op count, or `errors > 0`. Postgres runs need a throwaway
-database via `--pg-url`; nothing here touches `public`. Run from `backend/`:
+`scripts/benchmark/concurrency/` compares SQLite and Postgres contention on
+`users` with separate OS processes. `worker.py` connects via SQLAlchemy with
+the app's SQLite PRAGMAs after the orchestrator runs Alembic once.
+`run_concurrency_bench.py` uses a disposable schema outside `public` in the
+throwaway database supplied via `--pg-url`. It starts timing after a READY/GO
+barrier and fails on a crash, short op count, or any errors. Run from `backend/`.
 
 ```bash
 uv run python scripts/benchmark/concurrency/run_concurrency_bench.py \