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
AI Agent Reliability
12 articles connected to this topic.
MCP tool results need a materialization boundary
MCP tool results can include resource links, audio, images, and structured payloads. A materialization boundary keeps rich tool output from poisoning agent sessions.
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.
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.
Subagent workspace isolation: how OpenClaw keeps parallel agents from corrupting each other
Subagent workspace isolation gives each spawned agent its own working directory and prompt-local context, so parallel OpenClaw runs don't overwrite files or leak state. Here's how it works in v2026.5.28.
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.
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.