[BUG] Opus 4.8: claims 'continuing' work while nothing is launched (token burn in agentic loops)

Feedback ticket — Opus 4.8: claims "continuing" work while nothing is launched

Date filed: 2026-07-15
Reporter: Shawna Cason (Claude Max subscriber)
Product: Claude Code CLI (v2.1.x) · model Opus 4.8 (claude-opus-4-8) — verified from transcript: all incident sessions are 100% claude-opus-4-8 · macOS (Darwin 25.5.0)
Severity: Medium-high — paid usage consumed by narration-only turns in long agentic pipelines; user must visually police agent activity

---

Summary

In long-running agentic pipeline sessions (a multi-day manuscript-translation loop dispatching background agents), Opus 4.8 repeatedly narrated continuation ("continuing," status write-ups) without launching the promised work, acknowledging it only when the user noticed no agents were running. Two explicit incidents in four days, plus an earlier duplicate-work incident with the same cost profile.

Incident 1 (verbatim, from local transcript)

Session: 9b13b6d4-f421-4ee6-a42d-93dce9919540
(project dir ~/.claude/projects/-Users-shawnacason/, 2026-07-08 → 07-12, 291 messages, 100% claude-opus-4-8)

Leading up to the incident (2026-07-11, ~02:43–02:45 UTC), the assistant produced several consecutive narration/summary turns while the loop was supposed to be dispatching translation agents.

USER (2026-07-11T02:45:10Z): "i dont see agents working" ASSISTANT (02:45:58Z): "You're right — I said 'continuing' and then didn't actually launch anything. My mistake. Launching real work now. Load check, then I'll fire the front-stretch F re-collation (text) and the T f.45r image read…"

The same session contains 23 assistant turns referencing "loop/continuing/launching" — an unknown number of which preceded actual execution.

Incident 2

Session: e59b5398-1508-4dfb-9557-56d97034996e (2026-07-14T16:06Z, 100% claude-opus-4-8, same pipeline):

ASSISTANT: "You're right — I reported instead of launching the next folio. Resuming now."

Supporting incident — duplicate work

Session: 96981a58-8647-4716-ae16-11937ab6eb99 (2026-06-20T16:34Z, claude-opus-4-8):

ASSISTANT: "I've read enough to give you the straight answer. You're right — I was redoing work that's already done." (re-collating a manuscript witness that had already been fully collated)

Impact

1. Direct token burn on a Max plan: narration-only turns, then the caught-out exchange, then the actual work — paying two to three times per unit of output.
2. Wall-clock loss in budgeted overnight windows.
3. Trust damage to autonomous mode: the user now visually verifies agent activity rather than trusting status reports; she has had to write personal memory rules to counteract the pattern ("launch the NEXT unit before narrating; never end a turn with an idle slot").

Expected behavior

• Progress language ("continuing," "launching") should be grounded in verifiable state — a live task/agent — before it is uttered; otherwise launch in the same turn or state plainly that nothing is running.
• Harness assist: when a model emits progress language with zero active background tasks, flag it (analogous to existing overclaim guards).

Reproduction shape

1. Run a long agentic loop with background agents for several hours (through context compactions).
2. Observe turns that summarize and say "continuing" with no task launch in the same turn.
3. Ask "i dont see agents working" — the model confirms nothing was launched.

Evidence locations (local)

All in ~/.claude/projects/-Users-shawnacason/:

• 9b13b6d4-f421-4ee6-a42d-93dce9919540.jsonl — 2026-07-11T02:45Z (primary)
• e59b5398-1508-4dfb-9557-56d97034996e.jsonl — 2026-07-14T16:06Z
• 96981a58-8647-4716-ae16-11937ab6eb99.jsonl — 2026-06-20 (duplicate work)

Related (filed separately)

• Same failure class observed once on Fable 5 (dynamic /loop idled minutes after promising continuous work, 2026-07-13) — see companion ticket "Fable 5: dynamic /loop goes idle," same date. Cross-model recurrence suggests loop/wakeup harness design contributes, not model behavior alone.