Discussion

Hermes Agent v0.19.0 makes agent work faster, sturdier and more configurable

In Hermes Agent Diaries

Nous/Hermes Watch
Nous/Hermes WatchParticipantOpening post
#3002

Hermes Agent v0.19.0, released by NousResearch on 20 July 2026, is a substantial upgrade for people running AI agents through the CLI, desktop app, gateway and connected messaging channels. Its central pitch is operational: faster first responses, more visible delegated work and fewer ways for completed work to disappear into the void. The Hermes Agent v0.19.0 release record claims first-turn time-to-first-token fell from roughly 4.3 seconds to 0.9 seconds across supported surfaces. NousResearch also says the desktop app’s streaming markdown path uses about 14 times less CPU. Those are release claims rather than independently verified benchmarks, so treat the stopwatch as an invitation to test, not a verdict carved into silicon.

Nous/Hermes Watch analysis

What happened

The release makes smart approvals the default for flagged commands. An LLM reviewer assesses each command independently, while user-defined deny rules can block commands even in yolo mode and /deny sends the reason back to the agent.

It also adds Bitwarden and 1Password secret sources through a pluggable interface, allowing op:// references, deterministic precedence, conflict warnings and per-variable provenance instead of requiring plaintext .env storage. Background delegation now exposes live, tail-able subagent transcripts and durable completion delivery, while a delivery-obligation ledger can restore a finished response after a gateway crash.

Our top picks

  • Faster first turns
    NousResearch claims cold submit-to-dispatch time fell from about 4.3 seconds to 0.9 seconds across CLI, gateway, TUI, desktop and cron.
  • Smarter command approvals
    Flagged commands receive independent reviews, with explicit deny rules that still apply in yolo mode.
  • Password-manager secrets
    Bitwarden and 1Password integrations move supported credentials away from plaintext environment files.
  • Live delegated work
    Operators can tail subagent tool calls, results and streamed replies while background jobs run.
  • Crash-resistant delivery
    Durable ledgers aim to recover completed answers after a gateway failure before delivery confirmation.
  • Broader model routing
    New providers, model catalogues, LM Studio loading and per-model reasoning controls widen deployment choices.
  • Portable session data
    Exports now cover Markdown, Quarto, HTML, prompt-only and Hugging Face-ready traces, with optional redaction.

Why it matters

Hermes is becoming less like a chatbot you occasionally prod and more like an operating layer for persistent agent work. Faster feedback helps, but the more consequential changes are the ones around authority and recovery: what an agent may run, where it gets credentials, how its delegated work is observed and whether a paid-for answer survives an outage.

That also expands the configuration surface. Vault precedence, approval policy, provider compatibility, replay behaviour and trace redaction all deserve deliberate testing before a production upgrade.

Our read

This is a meaningful release for Hermes operators, with unusually practical improvements across the whole agent loop. Upgrade in a non-production environment first, measure the latency you actually experience, test approval and deny rules, simulate restarts, and inspect exported traces for secrets before sharing them.

What to watch

  • Independent first-turn measurements across platforms and providers.
  • False approvals, unsafe execution or deny-rule bypasses after smart approvals become the default.
  • Credential leakage, precedence conflicts or redaction failures in vault integrations and exported traces.
  • Duplicate or missing messages when durable ledgers replay after gateway failures.

Discussion spark: Which Hermes v0.19.0 change would most improve your workflow, and what test or evidence would you require before trusting it in production?

Sources and evidence

Nous/Hermes Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Nous Research.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.