Hermes Agent 0.19 (the Quicksilver release, shipped July 20, 2026) is the first version where running multiple AI agents in parallel stops being a demo and starts being a workflow you can trust. The headline change is an ~80% cut in time-to-first-token — cold start dropped from ~4.3 seconds to ~0.9 seconds across every surface (CLI, gateway, TUI, desktop, cron) according to Nous Research's release notes. But the features that actually make multi-agent work practical are the ones you don't see in a benchmark: live reasoning streams, smart approvals by default, a durable delivery ledger, and per-channel profile routing. Together they solve the three problems that killed parallel agent setups before: you can't see what agents are doing, work vanishes when the gateway crashes, and background missions stall waiting for you to tap "approve" at 2 a.m.
This article breaks down what changed in v0.19, why each change matters for anyone orchestrating more than one agent, and the five-step setup that turns a single Hermes installation into a reliable multi-agent team without the usual babysitting tax.
TL;DR — Last verified: 2026-07-23
- Hermes Agent 0.19.0 ("Quicksilver") shipped July 20, 2026 — ~1,065 merged PRs, ~3,300 issues closed, 450+ contributors (source).
- Cold start dropped from ~4.3 s to ~0.9 s (~80% faster) across CLI, gateway, TUI, desktop, and cron. Vendor-reported, not independently benchmarked.
- New: live reasoning streams, smart approvals by default, Bitwarden/1Password secrets, live subagent transcripts, durable delivery ledger, profile-based routing, terminal subscription management.
- The bigger unlock is operational: these features make fanning out work to multiple agents safe enough to run unattended.
- Pricing/feature changes are volatile — re-check before relying on specifics.
What Is Hermes Agent 0.19 and Why Does It Matter?
Hermes Agent 0.19 is the latest release of Nous Research's open-source, MIT-licensed autonomous agent platform, which runs from a terminal, a server, or a messaging gateway instead of being locked inside an IDE tab. The v0.19.0 tag (v2026.7.20) was published on GitHub on July 20, 2026 by Teknium (Nous Research), covering ~2,245 commits and ~1,065 merged pull requests since v0.18.0 (GitHub release). The release is called "Quicksilver" because — as the notes put it — "Hermes is the messenger god, and this window we made him move like it."
The reason it matters beyond the version number is that the changes target the operational friction that stops people from running agents in parallel at scale, not just raw speed. If you've ever set up a multi-agent system and then abandoned it because you couldn't see what the agents were doing, work disappeared on a crash, or you had to babysit approvals, v0.19 is the release that addresses all three. Hermes Agent reached 219k GitHub stars and 41.5k forks by this release (GitHub repo), making it one of the most-adopted open-source agent platforms — and 0.19 is the version most worth building on if you're orchestrating several agents at once rather than just chatting with one.
How Fast Is the Cold-Start Improvement?
The cold-start improvement cuts the delay between hitting enter and the model receiving your first message from roughly 4.3 seconds to roughly 0.9 seconds — an ~80% reduction that Nous reports applies identically to the CLI, the messaging gateway, the TUI, the desktop app, and scheduled cron runs (PR #59332, PR #59389). The savings come from removing three pieces of blocking work from the startup path: Discord capability detection moved off the critical path, the Ollama probe is skipped for non-Ollama providers, and agent-init blocking work was removed.
These are vendor-reported figures, not independently benchmarked. Nous measured them on their own setup; your numbers will vary with your provider latency, machine, and config. That said, the direction of the change is structural rather than tuned — the work removed from the critical path is work that was happening whether you used those features or not. A few seconds of dead air sounds cosmetic until you're firing off dozens of agent runs in sequence (as you do in a multi-agent workflow); then the stall quietly eats your afternoon. If you're curious how this fits into the broader landscape of agent harness evolution, see our guide to AI agent harness evolution in 2026.
What Are the New Speed and Visibility Features?
The speed and visibility changes in v0.19 make agents feel alive and keep you informed while they work:
| Feature | What it does | Why it matters for multi-agent work |
|---|---|---|
| Live reasoning streams | Reasoning models now stream their thinking token-by-token by default (display.show_reasoning ON) instead of showing a spinner for 30 seconds. |
You can catch a drifting plan in 5 seconds instead of waiting on a finished wrong answer — critical when you're overseeing several agents. |
| Per-token response painting | The response box paints per token, not per line. | Streaming feels responsive even across long generations. |
| Desktop app speed wave | 20+ targeted performance PRs; 14× less CPU in the markdown splitter, virtualized review-pane diffs, snappy session switching. | You can monitor many agent sessions without the desktop app choking on large transcripts. |
| TUI incremental markdown | The terminal UI renders markdown incrementally. | Long tool outputs stay readable in the terminal where you orchestrate. |
| Live subagent transcripts | delegate_task dispatches return transcript files you can tail -f immediately; every tool call, result, and streamed reply is logged per child. |
You can open a background job mid-run and read what it's actually doing — instead of guessing or waiting for the final summary. |
The live subagent transcripts are the quiet upgrade that makes fanning out work to multiple agents trustworthy. Before this, you delegated a task and waited blindly; a job that drifted wouldn't surface until it landed as finished-but-wrong. Now you can watch each agent work in real time and intervene early. For more on organizing agent work into something more durable than a chat window, see our article on organizing your AI work like an operating system in 2026.
How Do Smart Approvals Work and Why Are They Now Default?
Smart approvals are an LLM-based reviewer that independently assesses flagged commands instead of asking you to approve every action manually. In v0.19, this behavior is on by default: when a command gets flagged, a model reviews the exact command instance first and decides whether to proceed. Each verdict covers only that exact command — a later, similar-looking command is reviewed again rather than inheriting a blanket approval (PR #62661).
The release also adds user-defined deny rules and a /deny <reason> command. Deny rules block commands even under yolo mode (the mode that skips most confirmations), and /deny tells the agent why you refused so it can course-correct. The plugin pre_tool_call approve action can also escalate to a human gate.
This matters for multi-agent setups because background missions no longer stall at 2 a.m. waiting for you to tap approve. You get the safety net without babysitting the screen. The tradeoff — which AgentRiot's coverage correctly highlights — is that delegating first-pass review to a model is only useful if the command boundary stays narrow and the human can still set hard limits. If you're granting agents production access, layer this on top of the access-control patterns in our guide to securing autonomous agents that run without you watching.
How Do Bitwarden and 1Password Integration Work?
v0.19 introduces a pluggable SecretSource interface so API keys no longer need to live in plaintext .env files. Hermes can now fetch secrets from Bitwarden and 1Password (op:// references) at load time. Multiple vaults can be enabled simultaneously with deterministic precedence, conflict warnings, and per-variable provenance tracking. Future vault providers drop in as plugins.
This consolidates eleven competing community PRs into one orchestrated interface — a sign of how much demand there was for the feature. The practical value in a multi-agent setup: each profile (agent role) can pull only the secrets it needs from the vault, rather than every agent sharing one config file full of keys. If you're wiring agents into production systems, treat this as the key-management layer that complements the production-access security patterns we cover here.
What Is the Delivery-Obligation Ledger and How Does It Prevent Lost Work?
The delivery ledger is the feature that closes a silent-loss window most multi-agent operators have hit but few can name. When the gateway generates a finished response but crashes before the delivery confirmation lands on the channel (Telegram, Discord, Slack, etc.), that response used to vanish. In v0.19, final responses are recorded in a durable ledger inside state.db around the platform send. If the gateway dies between generation and delivery, the response is redelivered on the next boot (PR #67181).
This matters because in a multi-agent workflow, an agent's output often feeds the next step. If the output disappears on a crash, the downstream agent either waits forever or runs on nothing. The ledger means work your agent already did doesn't vanish — which is the difference between a pipeline you trust and one you keep re-checking.
What Is Profile-Based Message Routing?
Profile-based routing lets a single multiplexed gateway sharing one bot token route specific guilds, channels, or threads to completely different profiles — each with its own config, skills, memory, and secrets (PR #64835). A "profile" in Hermes is a fully isolated agent instance: different API keys, models, session histories, skills, and memory.
Before this, running different agent roles (say, a researcher, a writer, and an SEO analyst) required either separate bot tokens or a single shared config. Now one gateway can route #research to the researcher profile, #content to the writer, and #seo to the analyst, each with only the tools and secrets its role needs. A misconfigured profile can't take down the whole gateway (the release documentation notes a multiplex hardening wave 2 around this). The GATEWAY_MULTIPLEX_PROFILES override is available for advanced setups.
What Else Shipped in v0.19?
Beyond the highlights, Quicksilver adds a broad set of features (GitHub release notes):
- Terminal subscription management —
/subscriptionand/topupcommands so you can see your plan, remaining allowance, and preview upgrade costs without leaving the TUI. The desktop app got a matching billing settings tab. - New providers — Fireworks AI (first-class, with cost estimation), DeepInfra, and Upstage Solar. The model catalog added GPT-5.6 (Sol/Terra/Luna + Pro variants), grok-4.5 (GA),
moonshotai/kimi-k3(1M context, replacing kimi-k2.x),claude-fable-5/claude-sonnet-5,tencent/hy3(GA), LM Studio (JIT model loading), and GLM-5.2 with native reasoning_effort controls. - Reasoning effort tiers — new
maxandultraeffort levels for GPT-5.6 and Codex top tiers, plus per-model control. - Durable background delegation — background delegation completions are now durable; results are restored and delivered through an ownership-checked ledger if the process restarts mid-run.
- Per-provider
enabled: falseflag andexcluded_providersconfig — hide providers you don't use. - Bedrock catalog wave — real context-window probing and 1M-context rows for current-gen Claude and Fable models.
Pricing, model availability, and effort-tier names are volatile — verify on the Hermes Agent docs before relying on specifics.
How to Set Up a Multi-Agent Workflow With Hermes 0.19 — 5 Steps
The real unlock of v0.19 is that fanning out work to multiple agents becomes practical rather than just impressive. Here's the setup that makes it work — built from the features above and the patterns in the official multi-agent tutorial and Hermes profiles docs.
Step 1 — Create one profile per agent role
Each profile is a fully isolated agent: its own config, API keys, model, skills, memory, and sessions. Give each a narrow SOUL/system prompt so the agent stays in its lane.
hermes profile create "researcher" --clone
hermes profile create "writer" --clone
hermes profile create "seo-analyst" --clone
A researcher profile gets web search and browser tools; a writer profile gets file and publish tools; an SEO profile gets search and analytics tools. Don't give every agent every tool — access scope is the first line of defense, as we detail in our agentic development security framework.
Step 2 — Route channels to profiles with a single gateway
With profile-based routing, one gateway (one bot token) routes channels to different profiles. Set GATEWAY_MULTIPLEX_PROFILES and route each channel to its profile. One misconfigured profile can't take down the others.
Step 3 — Delegate with live transcripts so you can watch work happen
Use delegate_task to fan out subtasks. In v0.19, each dispatched child returns a transcript file you can tail -f immediately. Every tool call, result, and streamed reply is logged per child in one human-readable log. Open any background job mid-run and read what it's actually doing — catch drift in seconds rather than landing on a finished-but-wrong result.
Step 4 — Trust the delivery ledger for durability
If the gateway crashes after work is done but before delivery, the finished response is recorded in state.db and redelivered on the next boot. This closes the silent-loss window for Telegram, Discord, Slack, and every other channel. You no longer lose an agent's completed work to an infra blip.
Step 5 — Let smart approvals run unattended, with deny rules as hard limits
Smart approvals are on by default: an LLM reviewer assesses each flagged command independently. Background missions no longer stall at 2 a.m. waiting for you. Set deny rules for anything destructive (production DB drops, deletions, spending thresholds) — these block even in yolo mode. Use /deny <reason> when the agent course-corrects so it learns from the refusal.
What This Means for You
If you're a solo operator or small team that has been running one agent in a chat window, v0.19 is the version where running a team of agents stops being a weekend experiment and starts being daily infrastructure. The five features that make the difference — live transcripts, durable delivery, smart approvals, vault-backed secrets, and profile routing — each remove a specific reason you'd otherwise have to sit and watch the agents instead of doing other work. Start with two profiles (a researcher and a builder), wire them through one gateway, delegate with live transcripts on, and add the deny rules for anything you wouldn't want a background agent to do at 2 a.m. The install isn't the hard part; the second day — deciding which profiles to create, which models to point at which station, and how to structure memory so agents don't contradict each other — is where the real work is.
FAQ
Q: Is Hermes Agent 0.19 free and open source? A: Yes. Hermes Agent is MIT-licensed and free to use. You pay only for the model API calls you make through it (via your own API keys or a Nous subscription). The repository is on GitHub.
Q: How much faster is Hermes Agent 0.19 compared to 0.18? A: Nous reports an ~80% cut in time-to-first-token on cold start, from ~4.3 seconds to ~0.9 seconds, across the CLI, gateway, TUI, desktop, and cron. These are vendor-reported figures; your actual numbers depend on your provider and setup. The improvement is structural (blocking work removed from the critical path) rather than tuned.
Q: Do smart approvals mean the agent can do anything without asking me?
A: No. Smart approvals mean an LLM reviewer assesses each flagged command independently instead of asking you every time. Each verdict covers only that exact command instance. You still set hard deny rules that block commands even in yolo mode, and you can use /deny <reason> to course-correct. The agent still escalates to a human gate for anything your deny rules catch.
Q: Can I run multiple Hermes agents with different models and API keys? A: Yes — that's what profiles are for. Each profile is a fully isolated agent with its own config, model, API keys, skills, memory, and sessions. With profile-based routing (new in 0.19), a single gateway can route different channels to different profiles. See the profiles documentation.
Q: What happens if the gateway crashes while an agent is working?
A: In 0.19, finished responses are recorded in a durable delivery ledger (state.db) before the platform send. If the gateway dies between generation and delivery, the response is redelivered on the next boot. Background delegation completions are also durable — results are restored and delivered through an ownership-checked ledger if the process restarts mid-run.
Q: Does Hermes Agent 0.19 work with Bitwarden or 1Password?
A: Yes. v0.19 adds a pluggable SecretSource interface that fetches secrets from Bitwarden and 1Password (op:// references) at load time. Multiple vaults can be enabled simultaneously with deterministic precedence, conflict warnings, and per-variable provenance. API keys no longer need to live in plaintext .env files.
Q: How do I update to Hermes Agent 0.19?
A: Run hermes update for existing installs, or pip install -U hermes-agent for pip-based installs. The release tag is v2026.7.20. See the GitHub release page for the full changelog.
Discussion
0 comments