Your agent works while you sleep.

How scheduled tasks work on Hermes OS

You define a task in natural language or structured format in the Hermes OS dashboard. Give it a name, write the task instruction — what the agent should do, what it should produce, and where the result should go — and set the schedule. The schedule accepts natural language ("every weekday at 8am") or standard cron expressions for precise timing. You can also pause, resume, or edit a task's schedule mid-run without losing its task history.

At the scheduled time, the Hermes OS orchestration layer wakes the agent, loads the relevant memory context, and starts the task. The agent runs through it autonomously and logs the result — output text, files generated, actions taken — to the task history. If the task produces a report or notification, it sends it to the configured destination: your Telegram, email, Slack, or a webhook endpoint.

Event hooks add conditional triggering on top of cron. Gateway hooks fire on every incoming and outgoing message — useful for logging, real-time alerting, and webhook forwarding to external monitoring. Plugin hooks intercept tool calls before and after execution, enabling metrics collection, guardrail checks, and custom business logic without modifying the agent's core configuration.

What runs well as a scheduled task

Anything repetitive and verifiable: the agent needs to be able to tell when it has done the task correctly. Competitive monitoring, daily summaries, data extraction from a defined set of sources, nightly code processing, weekly reporting — these all have clear success criteria and run reliably.

Tasks that are time-sensitive but not urgent in real time: a morning brief that arrives before you sit down is more useful than one that arrives while you are already buried in email. Scheduling it to run at 6:30am lets the agent do the research while you are sleeping and have the result ready.

Tasks that chain together: a monitoring task that triggers a follow-up research task when a condition is met, or a data extraction task whose output is automatically passed to a formatting task for the final report. These multi-step scheduled workflows replace what would otherwise be manual pipe-and-process scripts.

Examples running in the wild

Daily competitor brief: monitor 5 competitor pricing pages, check their changelog or blog RSS for product updates, and summarize any changes. Runs at 7am, arrives in Telegram as a formatted message by 7:15am.

Weekly codebase review: pull the last 7 days of merged PRs, check for patterns matching a defined list of technical debt indicators, and generate a markdown report linked to specific files. Runs every Monday at 6am, available for the standup.

Hourly error monitoring: check Sentry for new error classes, compare to a known-errors list, and alert via Telegram if a new error pattern appears. Runs every hour on the hour and sends a notification only when something genuinely new emerges.

Monthly lead enrichment cleanup: take all CRM contacts added in the past 30 days that still have empty company fields, look up each one from public sources, and update the records. Runs on the 1st of each month and produces a report of what it filled in.

Combining scheduling with persistent memory

When the same task runs repeatedly using the same agent profile, it compounds. A competitive monitoring agent that has been running for three months has a stored history of how competitor pricing has evolved. When it runs this week, it contextualizes the current pricing against that history rather than treating each observation in isolation.

This compounding effect is what separates a persistent scheduled agent from a cron script. The script runs the same logic every time. The agent improves its execution of the task as it builds up experience — faster navigation of familiar sites, better categorization of changes it has seen patterns of before, more relevant framing of results for your specific interests.