Hermes Agent v0.18.1 delegation and Vertex AI, OpenClaw mobile caching and GPT-5.6 support (agent tools for daily runs)

What shipped

Across this week from 6 to 13 July 2026 both Hermes Agent and OpenClaw released steady updates to their open agent frameworks. The changes center on parallel task handling, credential safety, mobile caching, and fresh model integrations. Builders now have more concrete options for running agents in chat, desktop, and production settings.

Hermes Agent

Hermes Agent delivered the bulk of this week's updates with 22 changes centered on v0.18.1. The release added background delegation, Vertex AI token handling, and project tools in the desktop app. These build on earlier versions and give users direct ways to run subagents or verify coding tasks without extra setup.

• •Background subagent delegation Hermes-agent now runs delegate_task to launch multiple subagents in parallel in the background and returns one consolidated summary. Users can handle complex jobs without blocking the main chat thread.
• •MCP security hardening The update locks down MCP config files, blocks cron credential leaks, redacts Slack tokens, and raises aiohttp version floors. Teams gain safer credential handling in shared environments.
• •Google Vertex AI integration Hermes-agent adds native Vertex AI support that mints and refreshes short-lived OAuth2 tokens from service accounts. Users avoid mid-session token expiry when calling Gemini models.
• •Gateway scale-to-zero support The gateway now idles to zero when unused and coordinates external drains before restarts. Production runs avoid dropped conversations during updates.
• •Desktop coding Projects A new sidebar, coding rail, review pane, and git worktree tools appear in the desktop app. Coders can organize tasks into agent-aware project folders.
• •Learn and journey commands The /learn command distills skills from folders or URLs while /journey shows an editable timeline of memories. Users build and prune reusable skills over time.
• •Completion contracts for goals /goal now uses contracts plus a verification ledger that runs real project checks. Agents must prove work instead of claiming completion.
• •Mixture-of-Agents presets MoA options appear in the model picker across CLI, TUI, desktop, and gateway. Users stream aggregated answers and inspect each model's reasoning.
• •Stability roll-up release Over 660 pull requests landed since v0.18.0, including Windows installer fixes and WhatsApp dashboard pairing. The package is more reliable on common platforms.
• •WhatsApp Baileys fix The dependency now uses the published npm release instead of a pinned git commit. Docker builds for the WhatsApp bridge succeed more often.
• •Hermes Agent v0.18.1 release Version 2026.7.7 ships the accumulated fixes and features from the prior cycle. Users can upgrade for the full set of delegation and security changes.
• •Hermes Agent v0.18.2 patch The follow-up release unpins the Baileys library and resolves remaining WhatsApp install issues. Docker users get cleaner image builds.
• •Multi-line prompt editor The /prompt command opens the system editor for drafting long markdown prompts. Terminal users can compose complex instructions without line limits.
• •Vertex AI Gemini connection Service-account JSONs now connect directly to Gemini models with automatic token refresh. Developers avoid manual token management in long sessions.
• •Production gateway scaling Scale-to-zero and drain coordination let teams run the gateway at varying loads. Active conversations stay intact during idle periods or restarts.
• •Desktop project management The app now stores coding tasks with a project-repo-lane model and worktree tools. Users keep agent work organized inside one interface.
• •Parallel background delegation Multiple subagents execute via delegate_task without freezing the chat and return a single summary on finish. Operators delegate fan-out work cleanly.
• •Journey timeline and graphs The /journey view and radial memory graphs let users edit or prune skills in CLI, TUI, or desktop. Memory management becomes visual and interactive.
• •Skill distillation via /learn The command pulls reusable patterns from directories, URLs, or past runs and writes them to project standards. Workflows become repeatable assets.
• •Verification evidence ledgers Coding tasks now require evidence from actual checks rather than model statements. Users confirm real completion before moving on.
• •MoA streaming with traces Individual model reasoning appears in labeled blocks before the final aggregator answer streams. Users see how each model contributed.
• •MoA provider in picker Named Mixture-of-Agents presets route prompts through ensembles automatically. The model picker surfaces these options in every interface.

OpenClaw

OpenClaw released 20 updates that emphasize mobile apps, new model families, and gateway session tools. The changes include offline caching, Telegram steering, and ClawRouter for dynamic routing. These extend the Claude Code-style agent to more surfaces and providers.

• •Meta muse-spark-1.1 support OpenClaw bundles the muse-spark-1.1 Responses API with streaming, tool calls, and encrypted replay. Users gain a new provider option with live validation.
• •Telegram Codex workflows Private /login pairs Codex while /steer and /tell commands guide runs. Final replies recover across common Telegram errors.
• •Openclaw attach command The new attach tool launches Claude Code against existing Gateway sessions with scoped MCP grants that revoke on exit. External harnesses connect safely.
• •OpenAI GPT-5.6 family Sol, Terra, and Luna models now work across API keys, OAuth, and Codex paths with correct context and pricing data. Users access the latest OpenAI releases.
• •Automatic session titles Utility models generate concise titles while Gateway groups add unread states, rename, fork, and archive controls. Chat history stays organized across devices.
• •Offline mobile caching iOS and Android apps pre-load bounded session caches and support spoken replies with on-device fallback. Apple Watch gains full voice turns.
• •Redesigned Control UI A searchable sidebar, context ring, and reasoning slider appear alongside a native macOS session browser. Navigation and model picking become faster.
• •ClawRouter provider Credential-scoped discovery and budget reporting ship for OpenAI-compatible and native Anthropic or Gemini transports. Routing adapts without manual config.
• •Crestodian onboarding Conversational agent-loop setup guides provider and model choices across CLI, web, and macOS with masked prompts. First-run friction drops.
• •ClawRouter plugin bundle The plugin adds dynamic model discovery and usage reporting for multiple transports. Budgets stay visible during runs.
• •Streamlined Control UI Session-first sidebar, compact meter, warm theme, and slash-command picker simplify active chats. Daily management takes fewer clicks.
• •Automated macOS Gateway The desktop app installs and starts a local Gateway automatically. Users skip manual setup steps on first launch.
• •iMessage poll support Native poll creation, voting, and usage footers appear in iMessage chats. Accounting shows per-turn costs inside the thread.
• •iOS and Android refresh iOS 26 visuals and expanded localization including Swedish reach both platforms. Flows for chat, talk, and reconnect improve.
• •Event-driven cron schedules On-exit schedules wake agents when watched commands finish. Runs detach cleanly without interrupting other workflows.
• •Telegram Codex pairing /login starts Codex sessions and /steer commands adjust them with recovery across API hiccups. Mobile steering becomes reliable.
• •Gateway attach for harnesses The attach command connects external tools to live Gateway sessions with temporary credentials. Testing stays scoped.
• •GPT-5.6 model family The catalog now lists GPT-5.6 models with capability metadata. Selection paths respect context and pricing correctly.
• •OpenClaw v2026.7.1-beta.6 The beta ships incremental fixes for the prior cycle. Early users can test the latest mobile and routing changes.
• •OpenClaw v2026.7.1-beta.5 This beta precedes the current one and contains earlier stability adjustments. Track the release notes for exact deltas.

What this means for you

For Vibe Builders: You can now run parallel subagents and distill skills with Hermes-agent commands while OpenClaw handles mobile sessions and auto titles. These updates let you ship agent workflows from desktop or phone without writing code. Test the /learn and attach commands this week to move from single chats to repeatable processes.

For Non-techies: Hermes Agent and OpenClaw now support background tasks, mobile voice replies, and automatic session titles for everyday use. Business owners can delegate work or check chats on iOS and Android without extra setup. Start with the desktop project tools or local Gateway to keep daily agent runs reliable.

For Developers: Production features such as scale-to-zero gateways, completion ledgers, ClawRouter discovery, and drain coordination landed in both frameworks. Compare Hermes-agent verification against OpenClaw attach grants before adding them to existing stacks. Benchmark the new Vertex AI and GPT-5.6 paths against current providers for latency and cost.

What to watch next

Track follow-up releases for Hermes Agent v0.19 and OpenClaw beta.7 expected next week. Watch for expanded Telegram and iMessage controls plus any new Mixture-of-Agents presets. Test one delegation or attach flow in a non-critical project before wider rollout.

Sources

• •[v2026.7.1-beta.6](https://github.com/openclaw/openclaw/releases/tag/v2026.7.1-beta.6)
• •[v2026.7.1-beta.5](https://github.com/openclaw/openclaw/releases/tag/v2026.7.1-beta.5)