Hermes Agent Bot Mode and OpenClaw 2.0 add group chats plus desktop controls
TL;DR
Hermes Agent and OpenClaw both shipped major updates this week that turn single agents into coordinated groups and give them direct control over browsers, desktops, and team workflows.
What shipped
Across this week Hermes Agent and OpenClaw released dozens of features focused on multi-agent coordination, desktop access, and safer team operations. The releases center on practical controls for running agents in daily tools rather than new model benchmarks.
Hermes Agent
Hermes Agent released version 2026.8.31 with Bot Mode as the core addition. The update adds multi-agent chats, peer messaging, and persistent memory for scheduled tasks while tightening security around instruction files and secrets. Desktop browser control and MCP server management round out the changes for users who run agents across Slack, Telegram, and local machines.
- •Slack and Telegram streaming Hermes now sends live progress cards into Slack channels and adds an inline command picker inside Telegram so users can trigger skills without leaving the chat app.
- •Write approvals for instruction files Changes to AGENTS.md, skills, and memory stores now require explicit approval to block prompt injection attacks before they reach the agent.
- •CLI command palette and metrics A Ctrl+P fuzzy search opens in the terminal along with a live status bar that shows latency and tokens per second during runs.
- •New model providers added Meta Model API, CommandCode, Tencent TokenPlan, and Nebius join the catalog with support for GLM-5.3-Flash, Gemini 3.7 Flash, and several other models.
- •Desktop browser navigation Agents can click, scroll, and read pages inside the built-in desktop browser tab or pop them out to the system browser for full context.
- •Unified MCP command center A single desktop screen now lists MCP servers with drag-and-drop import, health checks, and 30-day token estimates per server.
- •Mid-run subagent steering Users can watch child agents, send course corrections, or stop them early while still keeping any results already produced.
- •Persistent memory for cron jobs Scheduled runs now carry memory and prior outputs forward so agents keep context across multiple executions.
- •Hermes peer messaging Agents send direct messages to each other by handle across profiles and gateways with replies stored in durable bot chats.
- •Built-in Bot Mode group chats Desktop users can now open Discord-style rooms where multiple agent profiles join the same conversation using @-mentions.
- •Stronger secret redaction Terminal output and logs now strip secrets automatically while macOS TCC identity handling was improved for stability.
- •CLI status bar and pets Real-time cache and latency numbers appear in the terminal alongside a new /model filter and optional terminal pets.
- •Expanded provider list Meta Model API and Ramp Router were added with custom metadata overrides for the new models.
- •In-app browser control Agents gained direct click and inspect actions on the desktop browser tab with options to move pages to the system browser.
- •MCP server dashboard Drag-and-drop import and deep-link installs now sit inside one screen that also shows schema token estimates.
- •Live delegation controls The delegate_task command lists running children, validates JSON schemas, and tracks cost per delegation.
- •Cron continuity flags Jobs can set continuity=true to reuse prior outputs and skip LLM calls when a monitor detects no file changes.
- •Bot-to-bot messaging via CLI Agents exchange messages across gateways and store the full history in each bot's canonical chat for later review.
- •Multi-agent roster in desktop Named profiles with fixed avatars now appear in shared group chats inside the main app window.
- •Version 2026.8.31 release The update bundles Bot Mode so agents can operate as a coordinated group rather than isolated instances.
OpenClaw
OpenClaw released versions 2026.9.2, 2026.9.1, and 2026.8.2 plus a 2.0 milestone. The updates focus on faster chat interfaces, Mermaid diagram rendering, personal skill libraries, and the ability to run sessions on paired desktops or cloud workers. Team features such as operator roles and masked credential prompts were added for shared gateways.
- •Version 2026.9.2 release The update improves chat responsiveness and keeps dashboards and session state stable during long interactions.
- •macOS composer and voice updates The native app now offers a cleaner chat input, lower idle CPU use, and a persistent Talk overlay for voice.
- •Working directory configuration Users can set default or per-agent folders and run up to 100 concurrent Git worktrees under one root without session limits.
- •Anthropic Fable 5.1 support The CLI and dashboard now track usage for Fable 5.1 and keep GPT-5.6 Ultra selections across restarts.
- •Personal GitHub account linking Users can attach their own GitHub account to submit pull requests under a separate identity without re-entering credentials.
- •Android app UI alignment Chat screens and the composer now match the web Control UI with expandable input up to six lines.
- •Approval cards in chat channels Delegated requests now appear as interactive cards inside the original Telegram topic or Slack channel.
- •CLI update rollback Failed updates automatically revert and hand diagnostics to a triage agent while keeping secrets intact.
- •Personal skill libraries Users can import ZIP skills and publish them per identity on shared gateways without mixing with team tools.
- •Quick-start setup lane Fresh installs detect existing Claude Code or Codex keys and open the dashboard immediately.
- •Mermaid diagram rendering Diagrams now display interactively in web, macOS, iOS, and Android chats with enlarge previews on mobile.
- •Managed worktree support Agents can run 100 checkouts under a global root while respecting per-agent directory settings.
- •SuperGrok usage tracking Real-time token counts for SuperGrok appear in both CLI and dashboard panels.
- •GitHub identity separation Personal accounts can be added in the profile screen for distinct pull request authorship.
- •Android navigation match Sidebar and appearance settings now follow the Control UI layout exactly.
- •Channel-delivered approvals System agent proposals reach the requester's chat instead of staying hidden.
- •Gateway startup resilience The service isolates bad cron rows and keeps running on Windows after agent restarts.
- •Automated rollback on failure npm updates revert if Doctor checks fail and preserve prior configuration files.
- •Isolated skill sets on teams Personal libraries stay separate from workspace skills even on shared gateways.
- •Live credential detection New installs verify keys on the spot and skip the full wizard.
- •Cross-platform diagram previews Mobile apps retry failed Mermaid renders and offer larger views.
- •Version 2026.9.1 release The update adds diagram support to every chat interface.
- •Migration cleanup tools A dry-run command shows retained files so users can remove them safely after the Gateway stops.
- •iOS composer controls Inline model and permission settings now match the web version for queued messages.
- •Voice output filtering Internal reasoning is removed from spoken replies while tool audio stays intact.
- •Four new Control UI themes CRT, Manuscript, Rosé, and Miami options save offline and load without flash.
- •On-demand Chrome CDP relay macOS and Linux extensions can start a local relay even when the Gateway is offline.
- •Background session creation New sessions can launch without leaving the current page and keep placement settings.
- •Linux desktop companion .deb and AppImage builds now run from the system tray with keyboard shortcuts.
- •Home agent dock Cmd/Ctrl+Shift+H opens the Home agent beside the current view for quick context attachment.
- •Version 2026.8.2 release The update introduces the dockable Home agent for side-by-side work.
- •Team operator roles Named roles with scoped access and a shared SQLite credential store are now available.
- •Paired and cloud desktop control Agents can drive macOS, Windows, and Linux machines through UI actions or background calls.
- •Codex route migration Legacy codex references move automatically to standard openai paths via doctor.
- •OpenProse plugin removal The bundled plugin and /prose command are gone; doctor --fix cleans remaining config.
- •Masked credential prompts Agents request secrets through hidden fields that never appear in chat context.
- •Interactive widgets in chat Users can embed and pin widgets to dashboards with action permissions.
- •Structured question responses Agents receive answers via cards, buttons, or plain text with an explicit skip option.
- •Durable progress cards Session cards survive reloads and track subagent edits across all interfaces.
- •Sessions on paired devices Tasks run on local or cloud machines with automatic workspace sync.
- •Conversation search Exact word or phrase search now returns matching messages for quick reopen.
- •Version 2.0 release The largest update to date includes cloud sessions, real-time collaboration, and a rebuilt browser app after more than 16,000 pull requests.
- •SQLite team credential store Write-only secrets and environment variables are managed per team role.
- •Vision-based desktop control Paired machines are driven with UI elements and vision models.
- •OpenProse removal and route cleanup Legacy plugin and codex paths are replaced by upstream skills and openai routes.
- •Native media playback Video uploads and playback controls now work on iOS and Android across message threads.
- •Proxy-protected credential requests An opt-in proxy limits secret use to approved destinations only.
- •Dashboard widget export Pinned widgets can be exported as PNG images from any session view.
- •Interactive answer cards Structured questions support free-text alternatives and skip paths on every platform.
- •Cross-device progress tracking Durable cards follow subagent activity on web, desktop, and mobile.
What this means for you
For Vibe Builders: You can now run multiple named agents in one chat room with Bot Mode and keep them on schedule with persistent memory across cron jobs. Desktop browser control and approval cards let you steer agents from Slack or Telegram without extra setup. Start with the quick-start lane in OpenClaw to connect existing keys and test group chats this week.
For Non-techies: For daily business tasks, agents now handle browser navigation and desktop actions directly while sending approval requests into your existing chat channels. Team members can share skills and credentials safely through roles and masked prompts. Try the new dockable Home agent in OpenClaw to keep an assistant visible beside your current work.
For Developers: Production setups gain personal skill libraries on shared gateways, worktree management for up to 100 checkouts, and automatic rollback on failed updates. Hermes peer messaging and OpenClaw widget dashboards provide durable cross-agent state that survives reloads. Benchmark the new MCP health checks and delegation cost tracking against your current orchestration layer before rolling out.
What to watch next
Watch for follow-up releases on MCP server stability and any new model additions from Meta or Anthropic. Check whether OpenClaw adds more cloud worker options or Hermes expands Bot Mode to additional gateways.
Harsh’s take
Both projects spent the week duplicating features across two release trains instead of shipping one clean integration point. The repeated browser and messaging items suggest the teams are racing to match each other rather than solving distinct user problems. Builders should pick the tool whose gateway model matches their existing chat setup and test one multi-agent workflow end to end this week before the next wave of overlapping updates arrives.
by Harsh Desai
More AI news
- Daily RoundupMercury 2.5 and Nex-N2.5 on OpenRouter, Meta Muse agent, and Vercel speed gains for agents
Model releases and agent tools expand options for builders while enterprise deals and routing improvements reduce friction for teams shipping AI products today.
- Daily RoundupHugging Face model wave, Fal H3 Turbo video, and Product Hunt AI agents
Hugging Face saw five models trend including text, image, video and speech tools while Fal released an upgraded text-to-video model and Product Hunt featured new agent-style apps for note-taking and local coding.
- Daily RoundupQwopus3.8 trends on Hugging Face, Ponytail launches, and Replicate image LoRAs drop
New multimodal models and agent tools arrived on Hugging Face and Replicate while OpenAI addressed an agent takeover incident and a no-code reuse tool hit Product Hunt.