In the space of five weeks this summer, Nous Research shipped three major releases of Hermes Agent, its open-source AI agent framework, and the project crossed 224,000 GitHub stars. The wave started with Quicksilver, a release devoted almost entirely to speed, and peaked with Herald, the biggest single release in the project's history, which finally delivered capabilities users had been requesting since 2025.
Hermes Agent is an agent harness, not a model. It gives any LLM memory, tools, scheduling, and a place to run, across a CLI, a desktop app, and more than twenty messaging surfaces that share one agent identity. The July to August release wave is a useful snapshot of how far that harness has come.
Quicksilver: the speed release (v0.19.0, July 20)
The headline number of v0.19.0 is the cold start. The "Initializing agent..." latency dropped from roughly 4.3 seconds to about 0.9 seconds, an 80% cut that applies uniformly across the CLI, gateway, TUI, desktop app, and scheduled cron jobs, rather than being specific to one surface. The gain came from removing blocking work during initialization and adding token-keyed disk caching for capability detection.
Speed was only one of three pillars. Safety behavior got smarter: smart approvals, where an independent LLM reviewer assesses each flagged command on its own merits before it reaches the user, became the default. A new /deny command lets you tell the agent why you refused, and user-defined deny rules block specific commands even under fully autonomous "yolo mode." Secrets handling matured in parallel, with native Bitwarden and 1Password integration through a pluggable SecretSource interface, a real improvement over pasting API keys into plaintext config files.
Model support also widened considerably: GPT-5.6 across its Sol, Terra, Luna, and Pro variants, Grok 4.5 at general availability, Moonshot's Kimi K3, Claude Fable 5, and Google's freshly cut-price Gemini 3.6 Flash, with Fireworks AI added as a first-class provider carrying cost estimation.
- ~80% cold-start reduction, from 4.3s to 0.9s, across every surface
- Smart approvals by default, with per-command LLM reviewer verdicts
- Native Bitwarden and 1Password vault integration for secrets
- Durable background delegation with an ownership-checked ledger
Herald: voice, A2A, and a desktop platform (v0.20.0, August 3)
v0.20.0, dubbed the Herald Release, is the biggest single release in the project's history: roughly 3,650 commits, about 1,400 merged pull requests, around 5,200 files changed, and 650+ contributors since the previous version. It turns the agent from a text-based assistant into something that speaks, coordinates with other agents over a standard protocol, pushes signed events to external systems, and backs its research with verifiable citations.
The four headline features are worth unpacking. Real-time conversational voice streams replies clause by clause, supports barge-in so you can interrupt mid-sentence, and pairs with on-device wake words, meaning no audio leaves your machine while listening. A bundled A2A v1.0 plugin lets Hermes discover, talk to, and be driven by other agent-to-agent compatible agents, closing issue #514, one of the oldest open feature requests in the repository. Signed outbound webhooks reverse the integration model: Hermes now pushes lifecycle events with HMAC signatures to any HTTP endpoint you register, no polling loop required. And a grounded-citations skill matches every quote against the actual page text, with a fact-checking mode that audits any document or claim you hand it.
The desktop app arguably shipped its own release in miniature. Artifact cards provide versioned, sandboxed live previews of generated HTML or apps in a right-rail viewer. A real plugin SDK arrives with Kanban as its founding plugin, plus a global-hotkey quick-entry window, SSH remote-backend mode, and a second 60fps performance wave. As the release notes put it: the desktop stopped being a chat client and started being a workbench.
CLI power users were not left out. New commands include !command to run a shell command without spending a model turn, /init to generate an AGENTS.md for a project, and hermes import-agent, which migrates an existing Claude Code or Codex CLI setup into Hermes in one command. Under the hood, tools learned to recover from their own failures: truncated terminal output spills to a file the agent can read back, patch detects already-applied edits, and the default tool-calling iteration limit jumped from 90 to 500 for long autonomous runs.
- Conversational voice with streaming TTS, barge-in, and local wake words
- A2A v1.0 plugin for agent-to-agent interoperability, closing issue #514
- Signed outbound webhooks plus grounded citations with fact-checking mode
- Desktop artifacts, plugin SDK, and SSH remote-backend mode
v0.20.1 patch and the road to v0.21.0
On August 13, the project tagged v0.20.1, a stabilization rollup of roughly 656 merged PRs and 1,444 commits touching 2,172 files, closing around 481 issues. The patch spans the desktop app, gateway platforms, installers, tool system, and provider catalogs, and gives downstream consumers a stable tag for Docker images and hosted deployments. Full curated release notes for the whole window ship with v0.21.0.
The release cadence itself is the story. May brought Velocity, June brought Surface, July brought Judgment, which closed every P0/P1 issue, and Quicksilver, and August delivered Herald. The strategy is deliberate: the agent's edges, platforms, protocols, voice, and verifiability keep expanding aggressively while the core stays narrow.
Hermes Agent remains MIT-licensed and free to run on your own hardware, with hosted tiers for those who prefer managed infrastructure. Voice wake-word detection runs on-device, and TTS/STT providers are configurable, with free local options such as Edge TTS or Whisper available. For developers watching the open agent space, the Herald wave is the clearest signal yet that the harness, not just the model underneath it, is where the competition now lives.
Comments