Hermes Agent changelog: what's new each month
Every stable Hermes Agent release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.
September 2026
5 releases: 2026.9.7 → 2026.9.24
September brought major stability improvements to Hermes Agent, focusing on database concurrency issues, state corruption handling, and multi-platform message rendering. The month saw fixes to critical wedging problems and graceful degradation for edge cases that previously crashed sessions.
Highlights
- Fixed database locking issues that caused 4–20 second stalls when opening state.db.
- Eliminated sticky database errors from WAL generation and concurrent writes that left the store permanently wedged.
- Made full-text-search failures graceful; corrupted search indexes no longer fail entire conversations.
- Prevented single malformed database rows from crashing sessions list, export, or insights views.
- Stopped profile database cross-contamination from Desktop launch and session search operations.
New
- Matrix now renders LaTeX.
- Signal renders markdown tables.
- WhatsApp replies to messages (platform support).
- Telegram bots enforce @mention requirement when botsrequiremention is enabled.
Improved
- Remote dashboard sessions no longer expire during refresh bursts.
- Long-lived processes no longer leak duplicate state.db writer handles in gateway, dashboard, Desktop backend, and ACP.
- Opening state.db skips acquiring write lock when no writes are needed.
- Queued-lane finals refused by the platform are now recorded and redelivered.
- Stalled WebSocket sends are handled more robustly.
- Full-text-search errors now degrade gracefully instead of failing the entire conversation.
Fixed
- OpenZFS deleted dentries and close() racing appendmessage no longer produce sticky DeletedWalGenerationError.
- Bare display key in config.yaml no longer crashes gateway turns.
- TEXT timestamp or 1e30 epoch values no longer crash sessions listing.
- Malformed marker JSON no longer crashes jsonextract operations.
- More than 999 IDs in a listing no longer causes crashes.
- Sessions no longer bind to or read another profile's database under HERMESHOME override race conditions.
August 2026
8 releases: 2026.8.3 → 2026.8.31
August brought major voice and collaboration features to Hermes, including streaming conversational voice with hands-free wake words, Agent-to-Agent communication, and Bot Mode for multi-agent coordination. The month also expanded platform reach and added powerful new orchestration and CLI tools.
Highlights
- Streaming voice mode with barge-in lets you interrupt Hermes mid-reply, and custom wake words enable hands-free control on-device.
- Voice messages now work across WhatsApp, Feishu, DingTalk, LINE, QQ, Photon, and Weixin with platform-aware replies.
- Hermes agents can now discover, talk to, and coordinate with other Agent-to-Agent compatible agents via A2A v1.0.
- Bot Mode makes your agent profiles into a society with persistent memory, cron scheduling, and bot-to-bot messaging.
- Research output now includes grounded citations with fact-checking so every claim points to a verifiable source.
New
- Wake words and hands-free control with on-device detection.
- Outbound webhooks push signed lifecycle events to your systems instead of requiring polling.
- Grounded citations skill backs research claims with verifiable sources.
- Desktop app renders artifacts as versioned cards with sandboxed live preview.
- Agent-to-Agent v1.0 protocol enables agent discovery and inter-agent communication.
- Bot Mode with deterministic avatar faces and bot-to-bot DMs between agents.
- Live orchestration for subagents: list running children, steer them mid-flight, or stop early.
- In-app browser now directly navigated and clicked by agents, with page pop-out support.
Improved
- Voice input works on every major messaging platform with transcription and auto-TTS replies.
- Desktop app became a plugin platform with artifact viewer and quick entry.
- Cron jobs now load and update persistent memory, carrying output between runs.
- MCP servers and catalog merged into one coherent page with health checks and re-auth nudges.
- CLI added fuzzy command palette (Ctrl+P), /model picker with filtering, and /status showing reasoning mode and context usage.
- Six new model providers added: Meta Model API, CommandCode, Tencent TokenPlan, Nebius Token Factory, Ramp Router, and Actual.
- CLI commands run instantly with !command without spending a model turn.
- /init generates or updates AGENTS.md by scanning your project; /diff shows staged, all, or session changes.
July 2026
5 releases: 2026.7.1 → 2026.7.30
July brought major performance improvements and new ways to work with Hermes, including faster cold starts, smarter approvals, and deeper integration with your development workflow. The agent now learns continuously and can manage complex multi-task delegations without blocking your work.
Highlights
- Mixture-of-Agents is now a selectable model option alongside others, with full visibility into each reference model's reasoning.
- Cold-start time dropped from 4.3 seconds to 0.9 seconds, with a 20+ performance optimization wave across the desktop app.
- Smart approvals now use independent LLM review instead of asking you to manually approve every flagged command.
- Background task delegation lets you dispatch multiple subagents without blocking your chat, with live transcript monitoring.
- Desktop app now has first-class Projects with sidebar codebase management, coding rail, review pane, and git worktree handling.
New
- Mixture-of-Agents presets show as selectable models under a dedicated moa provider.
- /learn command turns anything (directory, URL, description) into a reusable skill.
- /journey command displays a timeline of memories and skills Hermes has learned about you.
- Background fan-out delegation lets multiple subagents run in parallel with live transcript files you can monitor.
- Password manager integration with Bitwarden and 1Password support via pluggable SecretSource interface.
- Profile-based message routing allows one gateway to serve multiple profiles across different channels or guilds.
- Delivery-obligation ledger ensures responses cannot be lost between generation and platform confirmation.
- Agent now verifies its own work on coding projects by running them instead of claiming completion.
Improved
- Hermes renders each reference model's full output as a labelled block in MoA ensembles so you can see what each model thought.
- Desktop app markdown splitter reduced CPU cost by 14× for long replies.
- Review pane no longer freezes on giant diffs thanks to virtualization.
- Gateway can now go dormant when idle and shut down cleanly without dropping in-flight requests.
- Subscription and top-up management moved from billing website into /subscription and /topup terminal commands.
- Smart approvals now use independent LLM assessment of flagged commands instead of requiring manual approval.
- Subagent delegation now returns live transcript files you can tail in real time.
Fixed
- WhatsApp bridge dependency now installs from published npm release instead of a pinned git commit.
- Agent no longer loses answers if the gateway dies between generating a response and confirming delivery.
- Cold-start initialization time reduced from 4.3 seconds to 0.9 seconds.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Hermes Agent: pricing, features and alternatives · All changelogs