diff --git a/docs/now/2026-09-21-active-is-not-working.md b/docs/now/2026-09-21-active-is-not-working.md
new file mode 100644
index 000000000..f603fdd80
--- /dev/null
+++ b/docs/now/2026-09-21-active-is-not-working.md
@@ -0,0 +1,8 @@
+# NOW -- Active is not working (2026-09-21)
+
+## `tri swarm` reads lanes, throughput and the model ranking in one screen (Closes #4512)
+
+- For eighteen minutes the swarm read `active 20` and finished six turns. Every bee was waiting out a 30-second backoff: NVIDIA answered 1,369 requests in twenty minutes with 429 or 503. `active` counts bees started and not finished; it cannot tell working from waiting.
+- `tri swarm` prints `/queen/status` as one screen: lanes, dispatch counts, and the server's model ranking (BrowserOS #496) with each candidate's cost per 500-token step, success rate, speed and tool-call verdict. `tri swarm --since 600` samples twice and reports finished per hour, the number a lane or model change should be judged by.
+- Exits 2 when the status endpoint does not answer: a swarm that is down is the finding, not an error in the tool.
+- What this does NOT establish: that a finished turn did useful work, or whether a low success rate is 503 (model overloaded - switching models helps) or 429 (one key's rate limit - fewer lanes per key helps). The server log separates those; this page does not.
diff --git a/scripts/tri_loop/swarm.py b/scripts/tri_loop/swarm.py
new file mode 100755
index 000000000..68e574d74
--- /dev/null
+++ b/scripts/tri_loop/swarm.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+r"""tri swarm -- one screen of the live swarm: lanes, throughput, and which model the bees are on.
+
+WHY THIS EXISTS
+---------------
+On 2026-09-21 the swarm read `active 20` for eighteen minutes while it finished
+six turns. Every bee was asleep in a 30-second backoff: NVIDIA answered 1,369
+requests with 429 or 503. `active` is the number of bees that have started and
+not finished, and it cannot tell working from waiting. The only reading that
+can is finished-per-interval, and the model ranking the server now keeps
+(BrowserOS #496) says why: each candidate's success rate over the last fifteen
+minutes, its measured speed, and whether it can call a tool.
+
+This prints both. With --since N it samples twice, N seconds apart, and reports
+the finished delta as a rate, which is the number to judge a lane change by.
+
+WHAT THIS ESTABLISHES, AND WHAT IT DOES NOT
+-------------------------------------------
+Established: what /queen/status answered, and, with --since, how many
+dispatches finished between two readings.
+Not established: that a finished turn produced useful work (read the verdicts),
+or why a model's success rate is low - 503 is the model overloaded, 429 is one
+key's rate limit, and this page does not separate them. The server log does.
+"""
+import argparse
+import json
+import os
+import sys
+import time
+import urllib.request
+
+DEFAULT_URL = "https://trios-agent-server-production.up.railway.app/queen/status"
+
+
+def read(url: str) -> dict:
+    with urllib.request.urlopen(url, timeout=30) as response:
+        return json.load(response)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(prog="tri swarm", description=__doc__.split("\n")[0])
+    parser.add_argument("--url", default=os.environ.get("TRI_SWARM_STATUS_URL", DEFAULT_URL))
+    parser.add_argument("--since", type=int, default=0, metavar="SECONDS",
+                        help="sample twice, SECONDS apart, and report finished per hour")
+    parser.add_argument("--json", action="store_true")
+    args = parser.parse_args()
+
+    try:
+        first = read(args.url)
+    except Exception as error:  # the swarm not answering IS the finding
+        print(f"tri swarm: status did not answer: {error}", file=sys.stderr)
+        return 2
+
+    rate = None
+    last = first
+    if args.since > 0:
+        time.sleep(args.since)
+        try:
+            last = read(args.url)
+        except Exception as error:
+            print(f"tri swarm: second reading failed: {error}", file=sys.stderr)
+            return 2
+        done = (last.get("dispatches", {}).get("finished") or 0) - (
+            first.get("dispatches", {}).get("finished") or 0)
+        rate = done * 3600 / args.since
+
+    if args.json:
+        print(json.dumps({"status": last, "finishedPerHour": rate}, indent=1))
+        return 0
+
+    workers = last.get("workers", {})
+    dispatches = last.get("dispatches", {})
+    print(f"lanes      {workers.get('active')}/{workers.get('capacity')} active "
+          f"(active counts sleeping bees too; judge by finished/hour)")
+    print(f"dispatches total {dispatches.get('total')}  finished {dispatches.get('finished')}  "
+          f"unreviewed {dispatches.get('unreviewed')}")
+    if rate is not None:
+        print(f"throughput {rate:.1f} finished/hour over {args.since}s")
+    models = last.get("models")
+    if not models:
+        print("models     ranking off (TRIOS_QUEEN_WORKER_MODEL_CANDIDATES unset)")
+        return 0
+    print(f"model      {models.get('chosen')}")
+    for m in models.get("ranking", []):
+        score = m.get("score")
+        cost = f"{score:6.1f}s" if isinstance(score, (int, float)) else "     -"
+        flags = " GONE" if m.get("gone") else ""
+        print(f"  {cost}  ok {m.get('successRate'):.2f} n={m.get('samples'):<4} "
+              f"tps {str(m.get('tokensPerSecond') or '-'):>4} tools {str(m.get('toolCalls')):<5} "
+              f"{m.get('model')}{flags}")
+    print("cost = expected seconds per 500-token step; '-' = not yet measured or no tool call")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())