Anthropic extended thinking needs session recovery, not manual transcript surgery
Anthropic extended thinking can fail after cache expiry or Gateway restarts. OpenClaw's beta fix shows how agents should recover without deleting session history.
Tag archive
Self-Hosted AI
23 articles connected to this topic.
Interrupted tool calls are the recovery test for production AI agents
Interrupted tool calls expose whether an AI agent can preserve state, resume safely and report failure clearly instead of leaving a user-facing run half-finished.
AI agent timeouts: why provider requests need bounded failure paths
AI agent timeouts prevent stuck provider, plugin and tool calls from freezing a run; OpenClaw 2026.6.1 turns more wait states into bounded recovery.
Windows AI agent nodes are becoming first-class infrastructure
Windows AI agent hosting is moving from browser tabs to native nodes. Here is what OpenClaw 2026.6.1 changes for self-hosted Windows automation.
AI agent tool call errors: how OpenClaw recovers malformed tool arguments
AI agent tool call errors usually come from malformed arguments, not bad reasoning. Here is how OpenClaw v2026.5.28 repairs smart-quoted edit arrays and quarantines unsupported tool schemas instead of failing the turn.
AI agent rate limits: keeping scheduled jobs alive through 429s
AI agent rate limits break scheduled jobs when a 429 lands mid-run. Here's how retry budgets, exponential backoff with jitter, and preflight model fallback keep cron automations finishing instead of silently skipping.
Codex app-server reliability in 2026.5.27: how the runtime survives crashes
Codex app-server reliability is the quiet half of 2026.5.27. Here is how shared clients survive startup failures, hook relays live through restarts, and workspace memory routes through tools.
AI agent auth profiles: separate model logins without breaking agent runs
AI agent auth profiles separate model credentials by provider, runtime, and operator so self-hosted agents can migrate logins, recover cleanly, and reduce credential blast radius.
Meeting notes agent: why transcript-backed summaries matter for self-hosted AI
A meeting notes agent is only useful when its summaries trace back to clean transcripts, source chunks and replayable context. OpenClaw 2026.5.26 moves that path into the core runtime.
OpenClaw gateway performance in 2026.5.22: faster startup paths without hiding risk
OpenClaw gateway performance improved in 2026.5.22 through cached model metadata, leaner startup paths, locked npm packages, and sharper operator diagnostics.
On-device Android AI agent: what X-OmniClaw changes for personal assistants
On-device Android AI agent architectures like X-OmniClaw move perception, memory, and app control onto the phone. Here is what that means for self-hosted assistants.
xAI device code OAuth for headless AI agents
xAI device code OAuth lets headless AI agents authorize from SSH, containers, and remote hosts without a localhost browser callback or pasted API key.
OpenClaw plugin SDK: typed tool plugins for self-hosted agents
OpenClaw plugin SDK lets teams ship typed tool plugins without loading runtime code for discovery; this guide explains the build flow and where it fits beside MCP.
Agent provider plugins: why OpenClaw 2026.5.12 made installs leaner
Agent provider plugins keep self-hosted AI agents lighter by moving heavy channel and model dependencies out of the core install until operators need them.
AI agent tool policies: sender-scoped permissions for dangerous tools
AI agent tool policies should vary by sender, channel, and action risk. OpenClaw 2026.5.12 adds sender-scoped controls for safer self-hosted agents.
Provider plugins: how OpenClaw 2026.5.12 makes self-hosted agent installs leaner
Provider plugins in OpenClaw 2026.5.12 move heavy Slack, Bedrock, Anthropic Vertex, and sandbox dependency cones out of core installs so operators only pull what they use.
LLM idle watchdog: escalating silent model streams through profile rotation and model fallback
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: configuring backend failover for ACP turns
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: durable spooling for always-on AI agents
Telegram bot reliability improves when polling, queueing and reply delivery are isolated from the main agent loop instead of sharing one fragile event path.
Run a Discord voice agent with : permissions, capture, and STT tuning
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.
The /steer command in : redirect an active agent run without starting a new turn
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.
AI Agent Prompt Injection Is Now an Execution Boundary
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.
Browser agents vs chat agents: why the prompt-injection attack surface looks so different
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.