diff --git a/AGENTS.md b/AGENTS.md
index 5bd4939..6540493 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -13,7 +13,7 @@ Command Center and the Sentinel AI agent ship from **one repo, one image, one de
 | Code | `backend/app/` + `frontend/` | `backend/app/sentinel_agent/` |
 | Process group | `app` | `agent` |
 | Command | `uvicorn app.main:app` | `python -m app.sentinel_agent` |
-| Machine | 1 GB, always-on, owns the volume | 512 MB, scales to zero, no volume |
+| Machine | 1 GB, always-on, owns the volume | 512 MB, always-on, no volume |
 
 Both are the `sentinel-command` Fly app, built from the root `Dockerfile` and deployed by `.github/workflows/deploy.yml`. There is no separate agent app, agent image, agent workflow, or agent lockfile.
 
@@ -24,6 +24,8 @@ Four rules follow, and breaking any of them breaks a deploy:
 3. **`[processes]` overrides the Dockerfile `CMD`.** The `app` command in `fly.toml` must stay in sync with that `CMD`.
 4. **CI path filtering is asymmetric.** `push` is filtered (docs and Markdown only); `pull_request` is **never** filtered. `master` requires `Backend tests (sqlite)`, `Backend tests (postgres)` and `Frontend audit + build`, and GitHub reports *no status at all* for a workflow a path filter skipped — so a filtered PR trigger would hang every PR that missed it, presenting as a stuck check rather than a config error.
 
+The agent machine is kept **warm** (`min_machines_running = 1`) rather than scaled to zero. Fly's proxy waits only ~8s for an auto-started machine to bind its port, and this process needs ~10s (Python + the MCP SDK + Sentry + a deferred LiteLLM import) — so an auto-started machine was declared unreachable and the wakeup came back `RemoteDisconnected`. It was ~7s before LiteLLM, i.e. always marginal. ~$2/month buys the problem away; see the comment on `[[services]]` in `fly.toml`.
+
 The agent is a separate **process group** rather than a thread in the web app because a run holds base64 frames for up to 270s, and the segment cache is already budgeted 384 MiB of the web machine's 1 GiB. Sharing one machine is how the OOM killer takes every org's streams down at once. Being a separate *app* was never what bought that isolation.
 
 Self-hosting still works the same way: `python -m app.sentinel_agent` runs standalone with `AGENT_MODE=poll` and a per-org `osa_` key, needing no inbound connectivity. Agent docs are in `docs/SENTINEL_AGENT.md`. The code came from the `SourceBox-Sentinel` repo (archived 2026-09-09).
diff --git a/docs/SENTINEL_AGENT.md b/docs/SENTINEL_AGENT.md
index 15c4f14..3cfba62 100644
--- a/docs/SENTINEL_AGENT.md
+++ b/docs/SENTINEL_AGENT.md
@@ -12,7 +12,7 @@ notification fires              ┌─ POST /wakeup  (HMAC-signed)
                                          POST /api/sentinel/runs/{id}/start
                                          run agent loop (LLM ↔ MCP tools)
                                          POST /api/sentinel/runs/{id}/complete
-                                       return 200 — machine idles → auto-stops
+                                       return 200 — machine stays warm (see fly.toml)
 ```
 
 ## Architecture