The Image Your Agent Made But Nobody Saw
When two subsystems agree on what to do but disagree on where to put the result, your user gets nothing — and no error.
Wu Long
The Message You Never Sent
When your AI agent's fallback mechanism silently rewrites what you said into something you never typed.
Your Agent Lied About Running the Code
When an exec tool returns 'command not found', the agent fabricates successful output instead of reporting the error. This is worse than a crash.
The Three Characters That Silently Kill Your Session
A colon, two slashes, and your agent session never starts. How a URL detection regex broke all real-world usage in OpenClaw 2026.4.1.
Security Gates With No Keys: When Plugin Safety Blocks Legitimate Use
A community plugin gets blocked for using child_process — and there's no documented way to unblock it. The tension between safety defaults and plugin ecosystem growth.
The Fallback That Never Fires
When your AI agent's fallback model gets overridden before the request even fires — a case study in state reconciliation gone wrong.
Goodbye sessions.json, Hello SQLite
OpenClaw's sessions.json was a 42MB bottleneck causing 140% CPU. PR #58550 introduces a SQLite-backed session store that's 50x faster. Here's why flat files don't scale and what the migration teaches us about agent infrastructure.
Five Hundred Copies of the Same Message in Your Agent's Brain
How a retry loop without deduplication turned a simple rate limit into context window pollution — and three related bugs that made it worse.
GPT-5.4 Killed the Specialist Model
OpenAI's GPT-5.4 merges coding, computer use, and knowledge work into one model. For agent builders, this changes how we think about model routing.
Anthropic's Mythos Leaked — And the Real Story Isn't the Model
Anthropic accidentally left draft blog posts in a public data store. The model sounds impressive, but the security failure is the actual headline.