OpenClaw Floods Agents with Fixes as Hermes Stays Silent
TL;DR
OpenClaw delivered dozens of targeted updates to its coding agent across routing, memory, voice, and plugin tools while Hermes Agent issued no releases this week.
What shipped
From 18 to 25 May OpenClaw shipped a steady stream of practical improvements to its open-source coding agent. Hermes Agent recorded no activity. The changes focus on stability, provider integrations, and workspace controls that affect daily agent operation.
Hermes Agent
Hermes Agent: No releases this week from Nous Research's autonomous agent framework.
OpenClaw
OpenClaw issued the bulk of updates this week with targeted fixes across video routing, memory access, model handling, and voice sessions. The volume of changes centers on production reliability and plugin extensibility rather than new headline features.
- •OpenAI video routing fix Routes video edit requests to the correct /videos/edits endpoint instead of posting references elsewhere.
- •Public memory artifacts Exposes memory files through the active memory bridge so wiki imports work without internal dependencies.
- •Claude 4.x 1M context Moves 1M context support to generally available Claude 4.x models and drops the retired beta flag.
- •npm shrinkwrap for releases Adds generated shrinkwrap files to the root package and plugins to lock dependency graphs on install.
- •Row-level session helpers Gives plugins direct read and patch access to sessions without the legacy whole-store shape.
- •xAI Grok OAuth and search Reuses existing xAI auth for web search and adds model aliases plus timeout defaults.
- •Chat session picker search Adds search and load-more pagination so older conversations stay reachable without long initial loads.
- •Embedding providers contract Introduces a reusable embeddingProviders capability so embeddings work outside memory adapters.
- •External meeting-notes plugin Ships a source-only plugin with auto-start capture, transcript imports, and Discord voice as first source.
- •Gateway startup optimization Cuts model-listing latency from roughly 20 seconds to 5 milliseconds by reusing stable reads and lazy loading.
- •OpenClaw v2026.5.22 release Bundles the gateway performance and channel catalog improvements into the May 22 build.
- •Baileys WhatsApp upgrade Updates the Baileys library to 7.0.0-rc12 for better stability on WhatsApp channels.
- •MCP doctor warnings Doctor now alerts when sandbox policies hide configured Model Context Protocol server tools.
- •Task maintenance CLI details Tasks command outputs JSON explanations of retained and reconciled candidates.
- •OpenRouter routing policy Requests now respect provider-level routing policies unless overridden by model or agent settings.
- •Per-agent lean local model Allows the experimental localModelLean option on individual agents instead of globally.
- •Bundled Policy plugin Adds workspace conformance checks, doctor lint findings, and opt-in repair via CLI.
- •Discord voice profile context Voice sessions now include bounded IDENTITY, USER, and SOUL profile files by default.
- •Discord voice user following Sessions can follow configured users across channels with handoff and recovery handling.
- •Legacy skill allowlist removal Drops the old cat SKILL.md path so skills must load through the read tool.
- •OpenClaw v2026.5.20 release Ships the exec approval and skill loading changes in the May 20 build.
- •Telegram DM draft previews Shows native draft previews for tool progress while keeping final answers on the persistent path.
- •CLI tool plugin commands Adds defineToolPlugin plus build, validate, and initialize commands for typed plugins.
- •Browser modal dialog handling Surfaces pending dialogs in snapshots and lets users answer them via CLI.
- •Meme-maker skill Adds template search with local SVG or PNG rendering plus Imgflip hosted output.
- •Mac app settings redesign Applies consistent card layouts and cached navigation to permissions, voice, and debug panes.
- •Android Talk Mode realtime Switches to Gateway relay voice sessions with streaming audio and on-screen transcripts.
- •Codex harness upgrade Bumps the bundled Codex harness to 0.132.0 and refreshes model-list docs.
- •xAI device-code OAuth Enables remote and headless authorization for the xAI provider without a localhost callback.
- •Policy plugin conformance Performs channel checks and workspace repair through the CLI for consistent setups.
- •Discord voice follow sessions Supports multi-user handoff and DAVE recovery when following users into channels.
- •OpenClaw v2026.5.19 release Clarifies agent fix defaults around clean bounded refactors in the May 19 build.
- •OpenClaw v2026.5.18 release Includes the same agent fix guidance in the May 18 build.
What this means for you
For Vibe Builders: You can now pull in OpenClaw updates for voice sessions, meme generation, and meeting notes without writing code. The new Discord follow and Android realtime features let you test agent mobility on existing chat platforms. Watch the Policy plugin for workspace checks that reduce manual setup time.
For Non-techies: OpenClaw changes focus on making agent chats more reliable during voice calls and file handling. SMB owners gain easier transcript imports and draft previews on Telegram without extra configuration. The Policy plugin adds simple checks that keep workspace rules consistent day to day.
For Developers: OpenClaw added row-level session helpers, an embeddingProviders contract, and per-agent lean model flags that simplify production integrations. Gateway startup dropped to milliseconds and shrinkwrap now locks plugin dependencies. Evaluate the new MCP doctor warnings and task CLI details before rolling updates into existing agent runtimes.
What to watch next
Track the next OpenClaw stable after the beta series and any follow-up on the external meeting-notes plugin. Watch for Hermes Research activity that could re-enter the agent framework space.
Harsh’s take
OpenClaw's release cadence shows a team shipping many small reliability patches rather than large new capabilities. The volume of fixes around auth, memory, and voice suggests the core agent loop still needs hardening before it feels production-ready. Builders should test the Policy plugin and row-level helpers this week on a non-critical workspace to measure whether the changes reduce support tickets or simply add another configuration surface.
by Harsh Desai
Sources
- •[v2026.5.24-beta.2](https://github.com/openclaw/openclaw/releases/tag/v2026.5.24-beta.2)
- •[v2026.5.24-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.5.24-beta.1)
- •[v2026.5.22-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.5.22-beta.1)
- •[v2026.5.20-beta.2](https://github.com/openclaw/openclaw/releases/tag/v2026.5.20-beta.2)
- •[v2026.5.20-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.5.20-beta.1)
- •[v2026.5.19-beta.2](https://github.com/openclaw/openclaw/releases/tag/v2026.5.19-beta.2)
- •[v2026.5.19-alpha.1](https://github.com/openclaw/openclaw/releases/tag/v2026.5.19-alpha.1)
- •[v2026.5.19-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.5.19-beta.1)
- •[v2026.5.18-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.5.18-beta.1)
- •[v2026.5.16-beta.7](https://github.com/openclaw/openclaw/releases/tag/v2026.5.16-beta.7)
More AI news
- Daily RoundupCodex Gains Remote Control as Cursor and Lovable Push Agent Tools
OpenAI rolled out Codex remote access and spreadsheet integrations while Cursor added Jira automation and Lovable released a mobile app plus MCP server. Hugging Face models and Product Hunt launches filled the rest of the day with new audio, UI, and Mac automation options for builders and teams.