Video: "Hermes Agent V0.19 Just Changed AI Agents Forever!" by Julian Goldie on YouTube.

What Quicksilver actually means for response speed

The name is not marketing. Time-to-first-token — the gap between sending a prompt and seeing the first word appear — drops from 4.3 seconds to 0.9 seconds in v0.19. That is roughly an 80% reduction. Overall reply speed follows a similar pattern. Live thinking, which previously held back output until a full reasoning chain had run, now appears immediately as Hermes processes the task rather than making you wait for a blank screen to resolve.

In practice this changes the feel of working with the agent considerably. A 4-second wait feels like a system thinking. A sub-second response feels closer to a fast colleague. For anyone running multi-step tasks — long Oracle sessions, competitor monitoring sweeps, or code review chains — the cumulative difference across dozens of model calls in a session is significant.

Persistent jobs: the change that matters most for reliability

Previous versions of Hermes lost unfinished work if the application closed unexpectedly. v0.19 fixes this by writing every job and its final reply to a database. An agent running a long research task will now complete it and store the result even if you close the laptop, lose the connection, or restart the machine before the job finishes. When you reopen Hermes, the task has either completed or can resume from where it stopped.

This is paired with a delivery ledger that addresses silent failures — cases where an agent completed work but the output was never actually delivered to the user or the target system. The ledger records what was sent and confirms receipt, so silent-loss bugs that were previously hard to diagnose become visible and recoverable.

Smart approvals and security improvements

Smart approvals is now on by default. Rather than prompting for confirmation on every action, Hermes applies a risk-based filter: low-risk, routine actions proceed automatically; anything involving external services, file writes, or sensitive operations still requires a confirmation step. The result is fewer interruptions on standard tasks without reducing oversight on the things that actually warrant it.

v0.19 also adds Bitwarden and 1Password as supported secret sources. Instead of pasting API keys directly into Hermes configuration, you can point the agent at your existing password manager and it fetches credentials at runtime. This is a meaningful practical improvement for anyone running Hermes in a shared or semi-automated environment where leaving credentials in config files is a problem.

New models and multi-profile routing

The release adds four new models to the Hermes model picker: GPT-5.6, grok-4.5, Kimi K3 and Sonnet 5. Kimi K3 in particular is worth noting given the articles we have covered on it recently — it is free (or very cheap via the Moonshot API), open-source, and carries a 1-million-token context window that holds up well inside longer Hermes sessions. Swapping to it inside v0.19 is a one-command change that leaves all existing skills and memory intact.

Multi-profile routing is a new capability that lets you direct different tasks to different model configurations from within a single Hermes session. You might route fast, low-stakes tasks to a cheaper model and send reasoning-heavy work to a frontier model — all managed inside Hermes rather than manually switching between tool configurations.

Subagent monitoring and what is still missing

v0.19 adds live transcripts and monitoring for subagents — the worker processes Hermes spawns to handle individual steps within a larger task. Previously you could set a task running and see the final output without visibility into what each subagent was doing along the way. The new monitoring view shows each subagent's progress in real time, which helps identify where a task is stalling and why.

What v0.19 does not change: the quality of your outputs still depends on how well you have structured your goals and skills. A faster, more reliable agent running a poorly configured Oracle prompt will still produce generic content. The Quicksilver release is an infrastructure improvement — it makes the existing capabilities of Hermes faster and more dependable, not different in kind. If your setup was already working well, you will notice the improvement quickly. If it was not working, a faster loop just surfaces the problems more efficiently.

Where this connects to NordSys

Hermes v0.19 is a well-executed infrastructure release — the kind that makes a tool you already trusted substantially more pleasant to use. If you want an AI agent running inside your business without the configuration overhead, that is exactly what our AI Agents provide — named, briefed and managed for you, no setup fee, from £6 a day.

See our AI Agents →