Hermes Agent v0.20 voice and A2A protocol, OpenClaw mobile updates, and multi-agent tools to test now

What shipped

From 10 to 17 August the open agent space saw concentrated releases from two major projects. Hermes Agent delivered a series of v0.20 updates focused on real-time voice, agent-to-agent protocols, and desktop tooling. OpenClaw shipped mobile and security improvements that extend its Claude-style coding agent to more platforms and workflows.

Hermes Agent

Hermes Agent accounted for the bulk of activity with repeated v0.20 point releases that added voice, A2A protocol support, and desktop workbench features. The updates target both individual users running local agents and teams needing signed webhooks or grounded research output. Concrete capabilities such as mid-turn redirects and tool self-recovery differentiate these releases from earlier agent frameworks like Auto-GPT or LangGraph.

• •Hermes Agent v2026.8.16 release Hermes Agent v0.20.2 shipped as an incremental update for existing installs and focused on stability fixes after the larger v0.20.0 changes.
• •Conversational voice and A2A protocol Hermes Agent v0.20.0 introduced real-time voice streaming with barge-in, on-device wake words, the A2A v1.0 multi-agent protocol, and sandboxed desktop artifacts for live previews.
• •Desktop and tool stabilization Version 0.20.1 bundled roughly 656 pull requests to fix bugs across the desktop app, gateway platforms, installers, and provider catalogs for more reliable daily use.
• •Mid-turn redirects and self-recovery The agent now accepts course corrections during active turns, raises iteration limits to 500, and automatically recovers from truncated outputs or patch mismatches.
• •CLI power commands added New shortcuts let users run shell commands with an exclamation prefix, initialize projects, inspect context, view diffs, and import configs from Claude Code or Codex.
• •A2A v1.0 protocol plugin A bundled plugin implements the Agent-to-Agent standard so Hermes can discover and coordinate with other compatible agent systems without custom bridges.
• •Desktop workbench and plugin SDK The desktop app gained sandboxed artifact previews, a plugin SDK with Kanban support, multi-window layouts, and a global hotkey quick-entry window.
• •Signed outbound webhooks Hermes now pushes HMAC-signed lifecycle events for sessions, turns, and tool calls to external endpoints for CI and dashboard integration without polling.
• •Grounded citations skill A new skill produces research with exact source quotes and links while a fact-checking mode evaluates documents for verified and false claims.
• •Real-time voice streaming Clause-by-clause TTS with full-duplex barge-in, silence detection, and on-device wake words works across CLI, desktop, and gateway platforms.
• •v0.20.1 stabilization rollup The patch release incorporated 656 merged pull requests and 481 closed issues to stabilize the desktop app, gateways, and tool system.
• •Prompt caching and latency cuts Native caching for Anthropic and DeepSeek models dropped cold-start latency from 14 seconds to 1.8 seconds and removed deep copies in the UI.
• •Context compression overhaul Per-turn micro-compaction and proactive tool-result pruning keep long sessions coherent while protecting recent user messages from deletion.
• •Self-healing tools and steering Tools now auto-recover from common errors and users can redirect active turns without losing in-flight work or original context.
• •CLI migration utilities Instant shell execution, project scanning, context inspection, and one-command imports from other agent CLIs speed up terminal workflows.
• •A2A protocol integration The new plugin lets Hermes discover, interact with, and orchestrate agents from heterogeneous stacks using the v1.0 wire protocol.
• •Live artifacts and quick entry Sandboxed previews appear in a right-rail viewer while a global hotkey opens a compact entry window for fast commands.
• •Lifecycle event webhooks Signed HTTP posts for session and tool events enable real-time triggers in external automation systems without constant polling.
• •Verifiable research citations The grounded-citations skill links every claim to page text and adds a dedicated fact-checking mode for document review.
• •Full-duplex voice mode Streaming TTS supports mid-sentence interruptions and busy-aware silence detection across all Hermes interfaces.
• •v0.20.1 patch release The update rolled up 656 pull requests to fix bugs in the desktop app, gateways, installers, and provider catalogs.
• •Hermes Agent v2026.8.13 update Version 0.20.1 arrived for existing installs with the same stabilization focus as the prior point release.
• •Proactive context compaction Micro-compaction per turn and tool-result pruning maintain coherence while guarding against skill hallucinations after pruning.
• •Automatic tool error recovery Truncated outputs save to disk, patch edits detect mismatches, and file writes run instant disk verification.
• •Live turn redirection Users send corrections during execution and the agent adjusts course while keeping the original prompt and in-flight work intact.
• •Terminal workflow shortcuts The exclamation prefix runs shell commands instantly and new slash commands generate AGENTS.md files or show token usage.
• •External agent interoperability The A2A plugin enables discovery and command exchange with other standard-compliant agents in multi-agent setups.
• •Sandboxed artifact previews Generated code and HTML render live in a protected viewer while the plugin SDK adds Kanban and floating panes.
• •Signed webhook streams Outbound events carry HMAC signatures so external CI and home-automation systems can react without polling loops.
• •Grounded research mode Claims now include verifiable source quotes and a fact-checking mode reports verified, inaccurate, or unresolvable statements.
• •Open-vocabulary wake words Custom local wake phrases and multi-profile routing allow hands-free activation without cloud voice services.
• •Streaming TTS with barge-in Clause-by-clause speech output supports mid-sentence interruptions and full-duplex listening on every platform.
• •Approval policy automation Operators can build allowlists from history, apply consecutive-denial breakers, and require approvals for docker commands.
• •Context window optimizations Token threshold overrides and per-turn compaction keep long sessions responsive across different models.
• •Tool self-recovery upgrades Terminal, patch, search, and file tools now handle missing matches and whitespace issues without manual fixes.
• •Mid-execution course correction Guidance sent during a turn steers the agent while preserving original context and draft work.
• •Instant CLI shell access The exclamation prefix executes commands without consuming model turns and new slash commands manage context and diffs.
• •Heterogeneous agent connections The A2A plugin supports discovery and control across diverse agent software stacks.
• •Desktop plugin extensions Sandboxed previews and a plugin SDK with Kanban and multi-window support expand the desktop workbench.
• •Lifecycle webhook pushes Signed HTTP events for turns and tools feed external dashboards and automation without polling.
• •Verifiable source citations Research output now includes exact page quotes and a fact-checking mode flags false claims.
• •Conversational voice interface Streaming TTS with barge-in and on-device wake words works across CLI, desktop, and gateway platforms.
• •Per-turn context compaction Micro-compaction and proactive pruning maintain coherence while custom token thresholds prevent stalls.
• •Automatic failure recovery Tools diagnose truncation, whitespace mismatches, and zero-match searches without user intervention.
• •Active turn redirection Corrections sent mid-execution adjust the agent while retaining the original prompt and in-flight results.
• •CLI project management tools Instant shell commands, AGENTS.md generation, and context inspection commands streamline terminal sessions.
• •A2A multi-agent workflows The bundled plugin lets Hermes accept commands from and orchestrate other compliant agents.
• •Live artifact previews Sandboxed right-rail viewers display generated code and HTML while the plugin SDK adds custom UI panes.
• •Signed event delivery HMAC-signed POST requests for session and tool events integrate with external services without polling.
• •Grounded fact verification The new skill backs claims with page text and evaluates documents for verified or false statements.

OpenClaw

OpenClaw focused on mobile reach and operational safety with 23 updates that add Android agent switching, iOS offline caching, and provenance checks for plugins. These changes make the Claude-style coding agent more practical for users who switch devices or run agents in production environments. The releases also include new model support and secret-binding rules that reduce accidental data leaks.

• •Android live chat switcher Users can change the active agent from the live chat screen while preserving chat state, Talk mode, and the home canvas.
• •iOS offline chat caching Recent sessions and transcripts load from a bounded per-gateway cache so users can review conversations without a network connection.
• •Meta muse-spark-1.1 support The bundled model adds Responses API streaming, tool calling, encrypted reasoning replay, and onboarding flows.
• •Fish Audio speech synthesis Hosted S2.1 synthesis with streaming voice notes and telephony joins local Fish S2 Pro reference-voice streaming in macOS Talk.
• •Plugin provenance warnings Arbitrary executable plugins now require explicit --force acknowledgement while official ClawHub flows stay frictionless.
• •Isolated macOS app profiles Named instances keep separate state, preferences, Keychain, and Gateway services while host-global login remains shared.
• •SQLite snapshot backup CLI New openclaw backup commands create, verify, and restore compact global and per-agent database artifacts with fresh-target restore.
• •Shared channel plugin monitor A unified SDK monitor handles admission, polling, pruning, and handoff for IRC, Synology Chat, and Google Chat integrations.
• •GPT-5.6 Ultra model support Sol, Terra, and Luna variants work across OpenClaw and Codex engines with atomic runtime and thinking selection via slash commands.
• •Host-bound shared secrets Secrets in the shared store now bind to exact HTTPS destinations so unbound substitution fails closed before plaintext leaves.
• •Pinned MCP dashboard widgets Session MCP app views can be pinned as living widgets with revision-bound grants that protect tool interactivity.
• •Android agent switching The live chat screen now lets users switch agents without losing Talk mode or the anchored home canvas.
• •iOS offline transcript view Bounded per-gateway caches let users read recent sessions while offline and block sends until reconnection.
• •Fish Audio integration Streaming synthesis, voice notes, and telephony join local reference-voice streaming inside native macOS Talk.
• •Executable plugin checks CLI and chat installs from arbitrary sources require --force acknowledgement while catalog flows remain one-click.
• •macOS profile isolation Separate app instances keep distinct Keychain entries and Gateway services while global node services stay untouched.
• •SQLite backup and restore Compact verified database snapshots can be listed, verified, and restored to fresh targets via the CLI.
• •Channel plugin ingress monitor A shared SDK monitor manages durable admission and handoff for IRC and chat integrations.
• •GPT-5.6 runtime switching Atomic model, runtime, and thinking selection works across engines with fallback configurations via slash commands.
• •Destination-bound secrets Shared-store secrets tied to exact HTTPS hosts prevent accidental plaintext egress when bindings are missing.
• •Gateway CPU profiles Bounded three-node rig profiles captured during PR 124,528 testing provide performance signals for the release.
• •v2026.8.1-beta.2 release The beta tag marks the latest point release in the 2026.8 series with the mobile and security updates above.

What this means for you

For Vibe Builders: You can now run voice-enabled agents on desktop with Hermes Agent v0.20 and switch agents mid-chat on Android with OpenClaw. The A2A plugin and sandboxed artifacts let you connect multiple agents and preview outputs without writing code. Start with the desktop quick-entry hotkey or the Android live chat switcher to test multi-agent flows this week.

For Non-techies: Hermes Agent voice mode and OpenClaw mobile caching make AI agents usable on phones and laptops without setup friction. Signed webhooks and provenance checks reduce risk when connecting agents to calendars or automation. Try the on-device wake words or offline transcript view to keep daily tasks moving even when networks drop.

For Developers: Hermes Agent context compaction and tool self-recovery plus OpenClaw secret binding and plugin monitors give concrete signals for production reliability. Compare the 500-turn iteration limit and A2A interoperability against your current stack before adding either framework. Watch the next point releases for gateway scaling numbers and plugin SDK stability metrics.

What to watch next

Track the next Hermes Agent point release for expanded A2A plugin coverage and OpenClaw gateway scaling metrics. Monitor whether the gym waitlist incident triggers new safety benchmarks across both projects.

Sources

• •[pr-124,528-profiles](https://github.com/openclaw/openclaw/releases/tag/pr-124528-profiles)
• •[v2026.8.1-beta.2](https://github.com/openclaw/openclaw/releases/tag/v2026.8.1-beta.2)