Hermes Operations — Prompts, Delegation, Automation

Hermes Operations — Prompts, Delegation, Automation A workflow diagram generated by Archify. 01 / Inbound EX / Scheduled Triggers 03 / Voice 04 / Tool Execution 05 / Gateway 06 / Work Streams 07 / Outputs EX / Guardrails Intake Route Execute + deliver Cost + approval controls User Prompts · chat, text + voice · Inbound › Intake User Prompts chat, text + voice Cron Triggers · digest · watchdog · trials · Scheduled Triggers › Intake Cron Triggers digest · watchdog · trials Voice Stack · STT + TTS · Voice › Execute + deliver Voice Stack STT + TTS Tool Execution · shell, files, web · Tool Execution › Execute + deliver Tool Execution shell, files, web Prompt Gate · RUN or rewrite+approve · Gateway › Intake · Jev Prompt Gate RUN or rewrite+approve Jev Gateway · model + tool routing · Gateway › Route Gateway model + tool routing Delegated Coding · separate CLI agent · Work Streams › Execute + deliver · flash tier Delegated Coding separate CLI agent flash tier Scheduled Jobs · digest + watchdog · Work Streams › Execute + deliver Scheduled Jobs digest + watchdog HTML Artifacts · native deliverables · Outputs › Execute + deliver HTML Artifacts native deliverables Chat Reply · text + voice · Outputs › Execute + deliver Chat Reply text + voice Approval Checkpoints · risky + destructive · Guardrails › Cost + approval controls › Execute + deliver Approval Checkpoints risky + destructive Decision Logs · auditable trail · Guardrails › Cost + approval controls › Execute + deliver Decision Logs auditable trail record unattended, permissionless-safe verified results risky ops coding tasks cron fan-out act transcribe / speak digests pass / rewrite every prompt audio out Legend User UI Agent logic Policy Tool action Context / trace External system

Cost discipline

  • • Delegated agent: flash-tier default, model announced every run
  • • Jev gates: reflex decisions at ~$0.0000086/call
  • • Cron: only always-permitted operations, no approval prompts

Human in the loop

  • • Prompt rewrites shown and approved before acting
  • • Destructive / externally visible actions always confirmed
  • • Watchdog: silent when healthy, non-empty output = finding