A Scheduled Agent Is Only as Good as Its Failure Handling
The hard part of a scheduled agent is not the model call. It is making sure a failed run never looks like a quiet day. Lance Martin and CJ Avilla of Anthropic make that the center of their reference walkthrough for a daily brief built on Claude Managed Agents, currently in beta.
The agent reads Slack channels and GitHub pull requests, tracks what changed since its last run, and posts a short brief to Slack. The design has six parts: sources, a destination channel, a versioned agent configuration, a cron schedule, memory, and guardrails. Memory is split in two — a read-only preferences store the reader owns, and a read-write state store the agent owns. Guardrails include read-only tokens, a Slack allowlist and a per-run spending cap. Credentials stay in a vault outside the sandbox, which sees only a placeholder token swapped in for allowlisted hosts.
The failure handling is the useful part. Each source gets a bookmark rather than a fixed time window, so runs leave no gaps and no duplicates. A failed read keeps its bookmark and is reported as unreadable. Items are re-checked just before posting, and a post counts as sent only when Slack returns ok with a message timestamp. The authors suggest starting the spending cap at three to five times a normal run’s cost. The reference implementation sits in the claude-quickstarts repo under managed-agents/daily-brief.
The pattern scales past briefings. Anthropic’s sales team used Managed Agents to rebuild inbound handling and reported that escalated leads became opportunities more than twice as often as leads from the old contact form, and closed about five days faster. The build took weeks rather than months.
When an agent runs unattended, ship the verification with it: bookmarks, send confirmation, a cost ceiling.