Hermes Agent

Feature

Schedule Reliable AI Agent Work with Hermes Cron

Quick answer

Hermes cron runs fresh agent sessions or script-only tasks. Current jobs resolve a job model pin before the fleet default and main model, and wall-clock schedules use a job timezone pin or profile timezone. Verify execution, delivery and the actual artifact separately: queued output is not confirmed arrival. Human-dependent work should produce a draft for later attended approval.

Key Points

  • ✓Create one-shot, interval, or five-field cron schedules
  • ✓Run LLM-driven jobs or zero-token script-only watchdogs
  • ✓Pause, resume, edit, trigger, and remove jobs by ID or name
  • ✓Attach skills and restrict per-job toolsets
  • ✓Set an absolute workdir for repository context
  • ✓Block invalid providers, skills, and delivery targets before model spend with preflight
  • ✓Resolve the model from a job pin, cron fleet default, then the main model at fire time
  • ✓Deliver locally, to exact Slack channels or DMs, Telegram topics, and other platform targets
  • ✓Inspect immutable execution history with hermes cron runs
  • ✓Make supported reports continuable with session attachment
  • ✓Chain jobs with context_from while keeping sessions isolated
  • ✓Verify the real artifact or message instead of trusting schedule state
  • ✓Dashboard operations loop: inspect schedule, next run, logs, provider usage, and delivery target in Web UI, then require the real delivered report or artifact as proof.
  • ✓Distinguish scheduled, paused and completed lifecycle states from blocked_config run status
  • ✓Separate failed or unknown execution attempts from last_delivery_error
  • ✓Track recurring execution failure streaks without conflating delivery failures
  • ✓Use a job timezone pin or the profile timezone and inspect finite repeat exhaustion
  • ✓Pin per-job reasoning effort independently of interactive chat settings
  • ✓Bound agent inactivity, scripts, media sends, cleanup, and Bot Chat delivery separately
  • ✓Keep delivery_queued, delivery_failed and last_delivery_unverified separate before retrying

How It Works

  1. 1Create a schedule in chat, with /cron, or through the standalone CLI
  2. 2Hermes stores the schedule, runtime boundary, and delivery rule
  3. 3The gateway starts due work in a fresh isolated session or script-only process
  4. 4Hermes saves output and delivers it to the configured target
  5. 5You verify the report, alert, file, test, commit, or live URL

Real-World Use Cases

Decision-ready daily briefs

Collect current sources, rank what matters, and deliver a concise brief to an exact Slack channel, Telegram topic, Discord channel, or local output.

Quiet service watchdogs

Run a deterministic script every few minutes. Empty stdout stays silent; failures and meaningful alerts are delivered without an LLM call.

Verified publishing jobs

Update a page, run formatting and link checks, build, deploy, verify the live URL at phone and desktop widths, and report the evidence.

Repository maintenance

Use an absolute workdir so project instructions and file/terminal tools operate in the intended repository, then require tests and a clean diff.

Multi-stage pipelines

Connect collect, rank, and publish stages with context_from while keeping each run isolated, inspectable, and independently recoverable.

Under the Hood

The gateway daemon checks scheduled jobs every 60 seconds and runs due agent jobs in fresh sessions. Jobs are stored under the active Hermes home, output is retained locally, and a scheduler lock prevents duplicate ticks. Agent jobs use the cron platform's configured toolsets unless a per-job enabled_toolsets list narrows them. A workdir injects supported project instructions and points file, terminal, and code tools at an absolute existing directory; current workdir jobs bind directory state to each run's task identity and can use the normal parallel pool. Script-only no-agent jobs run files from ~/.hermes/scripts/ without touching a model: non-empty stdout is delivered, empty stdout is silent, and non-zero exits or timeouts alert. Hybrid jobs can use a pre-check script and emit {"wakeAgent": false} to skip an unnecessary model call. Current model resolution is a per-job pin, then cron.model / cron.model_provider, then the main model at fire time. Unpinned jobs can follow changed defaults; pinned jobs do not use the normal fallback chain. A pin is not a spend cap. Delivery can target local output, origin, exact chats/topics, configured home channels, or fan-out sets. Continuable delivery is opt-in, and context_from passes upstream output without pretending isolated sessions share memory.

Frequently asked questions

Which schedule formats does Hermes cron support?

Use relative delays such as 30m, recurring intervals such as every 2h, five-field cron expressions such as 0 9 * * 1-5, or ISO timestamps. You can also ask Hermes conversationally to create an appropriate schedule.

Can Hermes cron run without model tokens?

Yes. A script-only no-agent job runs a script under ~/.hermes/scripts/ and delivers its stdout directly. Empty stdout is a silent successful tick.

Do cron jobs inherit the current chat?

No. Agent jobs run in fresh sessions. Include every required path, source, constraint, output format, and success condition in the prompt or attached skills.

How do I prevent surprise provider spend?

Choose an explicit per-job model/provider pin when stable routing matters. Unpinned jobs follow the cron fleet default or main model at fire time; pinned jobs do not use normal fallback. Use provider-side budgets and bounded work because a model pin is not a monetary cap.

How do I verify a Hermes cron job really ran?

Use `hermes cron runs <job> --limit 20` to inspect claimed, running, completed, failed, or unknown attempts, then verify the actual message or artifact. Execution history does not replace destination proof.

What cron states should I check after a missed run?

Separate lifecycle (scheduled, paused, completed), attempt history and run status such as blocked_config. Then inspect delivery_queued, last_delivery_error, last_delivery_unverified and the exact artifact or destination.

Do delivery failures count toward the cron failure streak?

No. The recurring failure streak tracks execution failures. Delivery errors are kept separately so transport problems are not mistaken for failed agent work.

Can a scheduled task ask me before sending?

Make it save a draft and stop. Review that exact artifact in a later attended turn before authorizing a send. Continuable delivery can carry the brief into a reply but does not grant approval or guarantee delivery.

Related Features