[BUG] Severe Opus 4.8 quality/instruction-following degradation across BOTH Claude Code and Cowork (June 14–18, 2026) — agents ignore explicit instructions, act unprompted, draw conclusions from incomplete data

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Since ~June 14, 2026 I've had a sudden, severe, and SIMULTANEOUS quality drop across BOTH Claude Code (a multi-agent setup) and Claude Cowork — all on Opus 4.8. This is behavioral degradation, not just elevated request errors:

• agents ignore explicit, repeatedly stated instructions ("do not touch/read this file" — done anyway);
• they take actions I never authorized;
• they draw firm conclusions from incomplete/broken data (e.g. acting on a black, not-yet-rendered Unreal screenshot, in a loop);
• they contradict themselves within a single session;
• they repeat fixes that already failed.

Nothing on my side fixes it. It is not isolated to one agent or one session — it hit my Claude Code orchestrator, its subagents, the Cowork "assembler" project, AND the plain Cowork assistant, at the same time. I don't know if the root cause is the model, the serving stack, or the harness — but it is clearly not my configuration.

DETAILED TIMELINE

• Jun 12–13: team assembled, stable and productive (commits co-authored "Opus 4.8"). No issues.
• Jun 14: mostly productive; first warning sign in the evening — the Cowork "assembler" hung and needed a restart. (Status page: no incident logged Jun 14.)
• Jun 15: clear ONSET (midday→evening). The orchestrator took an Unreal viewport screenshot before the scene loaded, got a BLACK frame, "concluded" the scene was too dark, changed exposure, re-screenshotted → same black frame → looped, making real edits from invalid images. It also thrashed (posted "ready to test", then cancelled/held, repeatedly) and spun up a pointless 5-minute auto-commit "heartbeat" loop (~45 junk commits). (Status: Opus 4.8 elevated errors 06:20–08:56 UTC.)
• Jun 16: worse, and it SURVIVED remediation. I distilled all agent memory (orchestrator memory 650→48 lines), created archives, and added a strict "verify-before-claiming" rule injected into context on EVERY turn (hook-logged). Despite that, agents still ignored explicit prohibitions (I forbade touching the combat code; the orchestrator still had a subagent read it) and kept making unverified claims. The Cowork assembler degraded in parallel (contradicted itself, re-proposed non-working solutions). (Status: multiple incidents, incl. ~10% error rate across Sonnet/Opus 17:23–19:20 UTC.) I stopped ~21:00 local on the weekly limit.
• Jun 17: didn't use it (limit). (Status: FOUR separate Opus 4.8 incidents.)
• Jun 18 (today): resumed after the weekly reset. Mixed — on small, fresh single-task sessions the orchestrator behaved correctly; but the Cowork assistant (also Opus 4.8), in a fresh ~2% context session, made repeated reasoning errors today (confident conclusions contradicting facts I'd just given it; citing UI that doesn't exist; self-contradiction). (Status: a service disruption resolved 07:40 UTC.)

• Ruled out on my side: memory cleanup, repeated session restarts, rebuilt agent instructions,

and a per-turn "verify-before-claiming" guardrail — none helped. It hit 3 independent Claude Code
sessions AND Cowork at the same time with no config change → points to a model-serving / harness
regression rather than user config.

• Correlates with your status page: near-daily Opus 4.8 "elevated errors" incidents June 5–18

(incl. June 15 06:20–08:56 UTC; four on June 17; a disruption June 18). Note: those are "errors",
while what I'm reporting is QUALITY/behavior, which the status page doesn't track.

• Precedent: Anthropic's April 2026 public postmortem where backend/harness changes degraded

Claude Code quality for ~a month.

• Also reported via Claude Code /feedback — Feedback ID 8ab86ab3-8d92-4bcc-92d1-bb2224ebd787.
• Support conversations (Fin): 215474727535414, 215474754263344, 215474757264569.
• Contact: novembercog@gmail.com

• Claude Code: v2.1.181 (latest). Also affects Claude Cowork.
• Model: Claude Opus 4.8 (1M context), effort xhigh. Plan: Max 5x.
• OS: Windows (win32-x64).
• Checked clean on my side: /model = Opus 4.8, /effort = xhigh, /context ~2% (fresh sessions).

What Should Happen?

Expected: agents follow explicit instructions, don't act unprompted, and don't draw conclusions
from incomplete data — as they did on June 12–13.
Actual: the behaviors above, across both Claude Code and Cowork, since June 14.

Error Messages/Logs

Steps to Reproduce

Just use claude as allways

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude Code v2.1.181

Platform

Other

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

_No response_