Hermes Agent MoA presets and Projects, OpenClaw mobile apps (open agent tools this week)

What shipped

From 29 June to 6 July 2026, Nous Research and the OpenClaw team shipped repeated updates to autonomous agent frameworks. Hermes focused on orchestration, memory tools, and integrations. OpenClaw emphasized mobile access and messaging reliability.

Hermes Agent

Hermes Agent from Nous Research delivered dozens of updates across CLI, TUI, desktop, and gateway. Key additions include Mixture-of-Agents support with live reasoning visualization, first-class Projects for codebases, and commands like /learn for skills and /journey for memory management, compared to prior versions that lacked these agent orchestration tools.

• •Prompt editor command Hermes-agent added a /prompt command that opens the system editor for drafting complex multi-line markdown prompts instead of single-line input.
• •Vertex AI integration Hermes connected Gemini models via Google Cloud service accounts with automatic OAuth2 token minting and refreshing to avoid mid-session expiration.
• •Gateway scale-to-zero The gateway now supports going dormant when idle and coordinates external drains before restarts to keep active conversations intact.
• •Desktop Projects feature The desktop app introduced dedicated Projects with sidebar, coding rail, review pane, and git worktree management for organized coding workflows.
• •Parallel subagent delegation Hermes-agent runs multiple tasks with background subagents at once and returns one consolidated summary without blocking the chat.
• •Memory timeline and graph Users manage memories via the /journey timeline in CLI or TUI and an interactive radial memory graph in the desktop app.
• •Learn command for skills The /learn command automatically extracts reusable skills from directories, URLs, or workflows and formats them to match repository standards.
• •Completion contracts for verification Hermes records evidence for coding tasks and uses /goal contracts judged against test runs rather than model assertions.
• •MoA reasoning visualization The system streams labeled reasoning blocks from each reference model before the final synthesized answer appears live.
• •MoA presets as models Mixture-of-Agents presets are now selectable under a moa provider across CLI, TUI, desktop, and gateway interfaces.
• •Security hardening release Hermes secured MCP config persistence, blocked credential exfiltration, redacted tokens, and fixed CVEs in dependencies.
• •Atomic memory operations The memory tool supports an operations array for atomic add, replace, and remove edits to prevent budget issues.
• •Redesigned Skills Hub The hub now offers connected sources, featured sections, security scans, and previews before skill installation.
• •Web dashboard profiles Operators configure models, skills, and MCP servers visually with a global profile switcher without YAML edits.
• •Grok composer access Users reach the grok-composer-2.5-fast model via xAI OAuth with a 200k context window and no separate API key.
• •Automation Blueprints Natural language questions replace cron syntax for scheduling tasks across dashboard, CLI, TUI, and messenger.
• •Image-to-image editing The image_generate tool accepts an existing image plus prompt to modify, remove backgrounds, or convert sketches.
• •Async subagent dispatch delegate_task with background=true returns a handle immediately so users continue while tasks finish.
• •Desktop app upgrades The desktop app gained rebindable shortcuts, native notifications, live subagent windows, VS Code themes, and a resizable terminal.
• •Raft network adapter Hermes connects to the Raft agent network via a bundled platform adapter using a privacy-by-contract metadata bridge.

OpenClaw

OpenClaw released mobile apps for Android and iOS plus channel improvements for messaging platforms. Updates include scoped capability profiles, native iMessage polls, Telegram Codex steering, Slack relay mode, and support for the GPT-5.6 model family along with numerous stability fixes.

• •Scoped capability profiles OpenClaw prepares per-conversation tool and access boundaries without weakening the default profile.
• •Native iMessage polls iMessage now supports poll creation, reading, and voting with clearer per-turn usage footers in chat.
• •iOS visual refresh The iOS app adopted the iOS 26 visual system with improved navigation and expanded localization on Apple and Android.
• •Event-driven cron schedules A new on-exit schedule kind wakes agents when a watched command exits and supports clean session detachment.
• •Telegram Codex workflows Telegram starts Codex pairing with /login, steers active runs, and recovers replies after transient API failures.
• •External harness attach command The openclaw attach command launches harnesses against existing Gateway sessions for easier workflow inspection.
• •GPT-5.6 model family OpenClaw recognizes the GPT-5.6 family across catalog, capability, and runtime selection paths.
• •Automated dating script use-case One user built an Instagram outreach script with OpenClaw and Claude code that generated international DM responses.
• •WhatsApp group reply fix Group replies now correctly quote the user's triggering message instead of older bot messages.
• •Google Chat DM routing fix Newer direct messages are no longer misrouted as group conversations.
• •Telegram message recovery Automatic recovery resumes stuck Telegram messages after crashes or restarts.
• •OpenRouter catalog safety check Oversized catalogs are rejected before they exhaust memory and destabilize the gateway.
• •DeepSeek reasoning separation Final answers display separately from reasoning blocks for DeepSeek-style models.
• •Gemini 3.5 Flash full window Selection of Gemini 3.5 Flash now uses its full 1,048,576-token context without errors.
• •Telegram webhook resilience Direct and group messages continue during brief channel restarts and reloads.
• •Matrix E2EE memory leak fix A leak in Matrix E2EE gateways that caused gradual memory growth and crashes is resolved.
• •Reply consistency improvements Discord and Telegram replies stay attached to intended conversations across session changes.
• •Android and iOS launch The free open-source agentic program is now available on both mobile platforms.
• •Strict config guardrails OpenClaw fails closed without an interactive TTY and rejects empty TLS certificate paths.
• •Prompt cache stability Codex partial deltas and long-context prompt-cache improvements reduce lost progress during runs.

What this means for you

For Vibe Builders: You can now draft complex prompts in your system editor, delegate tasks to background subagents, and visualize memory graphs inside Hermes Agent without writing code. OpenClaw mobile apps and Slack relay mode let you run agents from your phone or messaging apps. Test the /learn command and Projects sidebar this week to ship agent workflows faster.

For Non-techies: Hermes Agent now handles multi-line prompts and skill extraction automatically while OpenClaw brings the same agents to Android and iOS with better iMessage and Telegram support. Set up Automation Blueprints or per-DM model overrides to automate daily tasks like scheduling or customer replies without technical setup.

For Developers: Hermes added atomic memory operations, completion contracts judged on test evidence, and MoA presets selectable as first-class models. OpenClaw introduced scoped capability profiles, RAFT wake bridges, and external harness attach for resuming Gateway sessions. Benchmark the new verification hooks and prompt-cache stability against your current agent stack before integrating.

What to watch next

Watch for Hermes Agent skill hub previews and OpenClaw Mattermost /oc_queue command adoption. Check next releases for further Raft network integrations and additional mobile channel controls.

Sources

• •[v2026.7.1-beta.2](https://github.com/openclaw/openclaw/releases/tag/v2026.7.1-beta.2)
• •[v2026.7.1-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.7.1-beta.1)