BACK TO ARCHIVE
2026-02-18P1 CRITICAL
CASE #0050

The Sonnet 4.6 Upgrade That Broke Everything

Anthropic released a model update and thousands of carefully configured OpenClaw agents stopped working overnight.

CONFIRMED
๐Ÿค– ROGUE BEHAVIOR๐ŸŒช๏ธ INFINITE LOOP
Incident Brief

When Anthropic released Sonnet 4.6, thousands of OpenClaw users discovered that their carefully tuned agent configurations were suddenly incompatible. Agents that had been running reliably for weeks began hallucinating, ignoring instructions, and entering infinite loops. The issue was subtle: the new model's longer context window and different instruction-following characteristics changed how it interpreted existing system prompts. Users who had optimized their prompts for Sonnet 4.5's behavior found those same prompts produced wildly different results with 4.6.

AFFECTED USERS: ~5,000

Root Cause

The Actual Culprit

OpenClaw auto-updated to the latest model without version pinning. Agent prompts that were tuned for one model's behavior became unreliable with the new version's different instruction-following characteristics.

What Was Done
[OK]Community shared workaround to pin model version
[OK]Users re-tuned their agent prompts for new model
[--]Blamed Anthropic for the change
[OK]OpenClaw added model version pinning in next release
Lessons Learned
๐Ÿ“Œ

Pin your model versions

Auto-upgrading to 'latest' means your agent's behavior is at the mercy of someone else's release schedule.

๐Ÿงช

Test before you trust

Model updates are not backward compatible for prompt behavior. Test your agents against new versions before switching.

๐Ÿ—๏ธ

Fragile prompts are technical debt

If a small change in model behavior breaks your agent, your prompts are too brittle and need structural improvement.

Comments (0)

Loading comments...

0/1000
Case Info
Case Number
#0050
Severity
๐Ÿ”ฅP1 CRITICAL
Severity Level
Date
2026-02-18
Affected Systems
โ€ข OpenClaw Agent Runtime
โ€ข Claude API Integration