Hermes X Roundup: June 27, 2026 - MoA 2.0 Crosses 4,500 Likes in 17 Hours, mr_r0b0t's /learn Trick Hits 22K Impressions, Hermes One Drops an npm Package

The day belonged to Nous Research and Teknium. The June 26 announcement that Hermes Agent now exposes MoA presets as virtual models crossed 4,500 likes, 409 retweets, 2,200 bookmarks, and 882,000 impressions inside 17 hours of posting - the largest single-day tweet in the Hermes timeline since the v0.17 release. Teknium's technical breakdown landed in parallel: 2,007 likes, 172 retweets, 863 bookmarks, 352,000 impressions, 132 replies. Two of the day's top three posts are the same announcement, told twice: once as positioning, once as engineering.

The headline numbers are the lift claims - 8% above Opus 4.8, 11% above GPT 5.5 on the upcoming HermesBench. The same announcement includes a video walkthrough and the structural argument: Mixture-of-Agents stops being a hidden tool flag and becomes a selectable model provider alongside OpenRouter, Nous, Vercel AI Gateway, and local catalogs. The previous-day deep-dive post covers the PR-level structure: 30+ commits in 72 hours, fusion.presets[] schema, picker registration across CLI / TUI / dashboard, one-shot /moa vs sticky /model selection. What the announcement tweet does that the deep-dive does not is set the positioning. "The strongest models are gated and access is granted only to a select few" is the marketing surface; the PR is the engineering surface. The same launch, framed twice for two audiences.

Teknium's parallel post is the engineering version of the same claim, with the same video and the same lift numbers but framed as an implementation pattern: "Combine any provider's models into a mixture of your own. Access your presets as if it were a normal model in Hermes." The reply chain turns on a real question - the upcoming HermesBench is, as of posting, not public, and the lift claims are self-reported. Teknium's responses confirm the methodology is "MoA using Opus and GPT" as references with the aggregator budget at 32K, but the dataset and full benchmark code are scheduled to ship with the benchmark release.

The third post of the day is a community one-liner that landed harder than the launch announcement's ratio would suggest. mr-r0b0t posted a one-line /learn prompt for the agent to ingest the Hermes Agent developer docs: 458 likes, 33 retweets, 754 bookmarks, 22,400 impressions, 26 replies. The bookmark-to-like ratio (1.65x) is the signal - this is a post the community is saving for later, not one they are reacting to in the moment. The 754 bookmarks on 22K impressions is the same engagement pattern the DeepSeek V4 thread produced yesterday: technical content that the builder audience flags for re-use.

The Spanish-language follow-up from Fer (ex Lacrimae Rerum) - "le dije a Hermes que se aprenda su documentación, lo hice el mismo día que salió /learn" - confirms the pattern from two days running: the /learn skill shipped recently, the community was already doing it with hand-rolled /skill calls, and the gap between the two is exactly one slash. mr-r0b0t's reply confirms he had been running a "Hermes self improvement" skill as a cron job before /learn shipped. Two independent builders arriving at the same workflow from the same direction, weeks before the upstream implementation: that is the /learn discovery signal, and the reply chain is the documentation.

The fourth post of the day is a fresh ecosystem build. Hermes One released the hermesone npm package - one command to install or update MCPs, skills, agents, and workflows from the Hermes Registry. 7 likes on the announcement itself, but the prior post about Hermes Registry (21 likes) and the underlying skill/agent registry community have been compounding all week. The release timing is also part of the MoA story: a registry for installable agent components is the deployment surface that MoA-as-virtual-model needs to be useful in production. The picker can register fusion/<slug> slugs; the registry can ship a MoA preset as a first-class installable.

Evan.Z closed the day with a Chinese-language breakdown of an architecture pattern for using Hermes with Apodex (via the OpenAI API) as an external reviewer on high-risk decision points - the same "智囊团" framing Anthropic's pseudo-correctness warning is built on. The parent article the tweet quotes (Lonely) crossed 117 likes, 51 replies, 19 retweets, 170 bookmarks, 87,000 impressions - the highest-engagement Chinese-language Hermes post this week. The pattern is the same one Jess @ FireTeam pushed on June 24 and Squadic shipped on June 26: Hermes is becoming the orchestration layer that other agents plug into, not just the agent that runs the prompt.

Notable Mentions

• Bolt posted a link-only update about Hermes Agent that hit 27 likes and 9 retweets inside 24 hours - the third-highest engagement of the day among single posts, all from the same builder audience that the MoA announcement pulled in.
• Prasenjit Sarkar called out the real memory problem in agent systems - "not what the model can recall, but how fast the context window fills with noise" - while pointing at Tencent's newly-open-sourced long-term memory system as the structural alternative.
• Sibyl Labs published a long-form X Article titled "Sibyl Labs: Advanced Infrastructure & Tools for AI Agent Workflows" that crossed 122 likes, 29 retweets, and 613,000 impressions in the window, the largest non-Nous-research post of the day.
• Elias Al broke down Tencent's open-source memory system: 61% token reduction and measurable task success lift, the second most-cited external post of the roundup.
• Kepler.1571 shipped ObserveCo as an official Hermes Agent observability plugin - real-time fleet health, circuit breakers, token usage, memory analysis (duplicates and contradictions), local dashboard, no cloud.
• Horbunov Dima reported independent MoA preset benchmarks matching the official claim: +8% over Opus 4.8, +11% over GPT 5.5 on the same benchmark, run on a Hermes instance configured against the open MoA provider.

The thread across the day is the model abstraction layer. MoA-as-virtual-model is the structural change: a Hermes user can now pick a fused model the same way they pick Opus, and the picker, the dashboard, the CLI, the TUI, and the registry all learned the new slugs at the same time. The mr-r0b0t /learn tip is the operational version of the same insight - a one-line prompt that teaches the agent the same set of abstractions. Hermes One's npm package is the installable version - one command wires any of the above into a working deployment. The day was not a single feature launch; it was the same feature launch landing on three different surfaces at once, each one reinforcing the others.