An LLM idle watchdog catches silent model streams before they hang an agent turn. The 2026.5.12 fix escalates through profile rotation and configured model fallback instead of leaving the turn stuck.
Agent runtime fallbacks let an ACP turn try a backup model or runtime when the primary is unavailable, so a single dead backend stops killing user-facing replies.
Telegram bot reliability improves when polling, queueing and reply delivery are isolated from the main agent loop instead of sharing one fragile event path.
How to run a self-hosted Discord voice agent on v2026.5.7: permission audit via channels capabilities, the new 2.5s capture silence grace default, and STT tuning that stops the bot interrupting people.
How /steer and active-run steering work in v2026.5.3+: send guidance to a running session without queuing a new turn, with worked examples and the safer queue defaults from v2026.4.29.
Microsoft's Semantic Kernel RCE research shows why prompt injection in AI agents is no longer just a text problem. Here's how self-hosted agent builders should think about tool boundaries.
OpenAI now says prompt injection may never be fully solved for browser agents like ChatGPT Atlas. Here is what that means for chat-channel agents and self-hosted setups.