Loop Daily: 2026-05-19
May 17 was a quiet day for the autoresearch / agentic loop conversation in English Twitter. Across keyword variants — autoresearch, agentic loop, agent loop, auto-research, autonomous research — only "self-improving agent" returned a real cluster, and most of that cluster is crypto-promo noise around AI-themed tokens. What's left after filtering is a small set of legitimate loop-relevant cases, but the signal that does emerge is coherent: Hermes Agent's xAI integration is the structural news, the Hermes-vs-OpenClaw head-to-head moved from rhetoric to actual benchmarks, and a few operators are still wiring multi-agent control planes for production work. Six cases below, plus the bigger-picture note that agent-loop discourse may be condensing into the operational-tooling side rather than the methodological-research side.
#1
@KSimback
https://x.com/KSimback/status/2055965846606839918
Built an agent poker engine that pits Hermes vs OpenClaw head-to-head. After 100 heads-up Texas Hold'em matches the score was exactly 50-50, neither agent decisively better out of the box. Mixed models across the 100 matches to surface trends, then ran a model-vs-model tournament: 8 models, best-of-7 series, each match to bankruptcy or 100 hands. Round 1 results: GPT-5.5 beat Qwen 3.6 4-0, Opus 4.7 beat GLM-5.1 4-1, Kimi K2.6 beat Grok 4.3 4-3 (the only upset, went full 7), Gemini 3.1 beat DeepSeek V4 4-2. This is one of the only public head-to-head agent benchmarks on real adversarial work — and it required someone building an agent engine themselves.
https://x.com/KSimback/status/2055965846606839918
Built an agent poker engine that pits Hermes vs OpenClaw head-to-head. After 100 heads-up Texas Hold'em matches the score was exactly 50-50, neither agent decisively better out of the box. Mixed models across the 100 matches to surface trends, then ran a model-vs-model tournament: 8 models, best-of-7 series, each match to bankruptcy or 100 hands. Round 1 results: GPT-5.5 beat Qwen 3.6 4-0, Opus 4.7 beat GLM-5.1 4-1, Kimi K2.6 beat Grok 4.3 4-3 (the only upset, went full 7), Gemini 3.1 beat DeepSeek V4 4-2. This is one of the only public head-to-head agent benchmarks on real adversarial work — and it required someone building an agent engine themselves.
#2
@grok
https://x.com/grok/status/2056136152462098846
Grok's official account published Hermes Agent setup notes: Hermes Agent is Nous Research's open-source, self-improving AI agent, runs persistently on a computer or VPS with long-term memory across sessions, and as a Grok subscriber you can connect your account via OAuth (no API key) to use Grok models inside it. The phrasing matters — a frontier-lab account directly promoting a third-party agentic harness is the opposite stance from Anthropic locking down `claude -p`. Connection point for the day's overall narrative: agentic loops are getting subsidized into existence by exactly the lab whose models they consume.
https://x.com/grok/status/2056136152462098846
Grok's official account published Hermes Agent setup notes: Hermes Agent is Nous Research's open-source, self-improving AI agent, runs persistently on a computer or VPS with long-term memory across sessions, and as a Grok subscriber you can connect your account via OAuth (no API key) to use Grok models inside it. The phrasing matters — a frontier-lab account directly promoting a third-party agentic harness is the opposite stance from Anthropic locking down `claude -p`. Connection point for the day's overall narrative: agentic loops are getting subsidized into existence by exactly the lab whose models they consume.
#3
@ghumare64
https://x.com/ghumare64/status/2055911069495767503
Hermes becomes the first agent where xAI permits feed scrolling without blockers — for X Premium subscribers. Concrete: research, video gen, image gen, chat etc, all addressable from inside the agent on a Premium subscription, no API spend. The "agents reading the feed" problem just got a sanctioned solution, where previously rate limiting and ToS made every attempt fragile.
https://x.com/ghumare64/status/2055911069495767503
Hermes becomes the first agent where xAI permits feed scrolling without blockers — for X Premium subscribers. Concrete: research, video gen, image gen, chat etc, all addressable from inside the agent on a Premium subscription, no API spend. The "agents reading the feed" problem just got a sanctioned solution, where previously rate limiting and ToS made every attempt fragile.
#4
@EveDotEth
https://x.com/EveDotEth/status/2055916619587625228
Multi-agent isn't a big-tech thing. The restaurant analogy: one chef for starters, one for mains, one for desserts, one head chef coordinating. Multi-agent = multiple AI models each doing one focused job with a coordinator on top. Why split? A single AI handling everything gets confused on complex tasks. Five focused agents passing results is where serious work gets done. CrewAI and LangGraph are how you actually build it — define each agent's role, tools, and handoff. The restaurant doesn't need one genius, it needs coordination.
https://x.com/EveDotEth/status/2055916619587625228
Multi-agent isn't a big-tech thing. The restaurant analogy: one chef for starters, one for mains, one for desserts, one head chef coordinating. Multi-agent = multiple AI models each doing one focused job with a coordinator on top. Why split? A single AI handling everything gets confused on complex tasks. Five focused agents passing results is where serious work gets done. CrewAI and LangGraph are how you actually build it — define each agent's role, tools, and handoff. The restaurant doesn't need one genius, it needs coordination.
#5
@QLyun35332
https://x.com/QLyun35332/status/2056083838749831469
Open-sourced an X search MCP that runs entirely on X Premium Plus quota. Wraps Hermes Agent's Grok-4.x x_search into 6 tools exposed via OAuth 2.1 — usable directly from Claude, Codex, or ChatGPT with zero incremental spend. This is the second time today an MCP shows up as the bridge: instead of asking Anthropic to add native X search, the community is wiring Hermes-as-a-search-tool into Claude Code via MCP. The path of least resistance keeps being "let one harness call another."
https://x.com/QLyun35332/status/2056083838749831469
Open-sourced an X search MCP that runs entirely on X Premium Plus quota. Wraps Hermes Agent's Grok-4.x x_search into 6 tools exposed via OAuth 2.1 — usable directly from Claude, Codex, or ChatGPT with zero incremental spend. This is the second time today an MCP shows up as the bridge: instead of asking Anthropic to add native X search, the community is wiring Hermes-as-a-search-tool into Claude Code via MCP. The path of least resistance keeps being "let one harness call another."
#6
@PwrAgentAI
https://x.com/PwrAgentAI/status/2055853703576289353
Open-source desktop coding agent paired with Telegram, Discord, Slack, Mattermost, Feishu/Lark, or LINE — start, resume, steer, and approve from any chat app. The pattern keeps recurring this week: the agent loop runs on the desktop, the human interface lives wherever the operator is already chatting. Mobile-first remote control of long-running agents is the missing primitive Anthropic hasn't shipped, and open source keeps shipping it instead.
https://x.com/PwrAgentAI/status/2055853703576289353
Open-source desktop coding agent paired with Telegram, Discord, Slack, Mattermost, Feishu/Lark, or LINE — start, resume, steer, and approve from any chat app. The pattern keeps recurring this week: the agent loop runs on the desktop, the human interface lives wherever the operator is already chatting. Mobile-first remote control of long-running agents is the missing primitive Anthropic hasn't shipped, and open source keeps shipping it instead.
📡 Eco Products Radar
Eco Products Radar
Hermes Agent (Nous Research) — central to today's signal: xAI partnership, Grok OAuth subscription path, poker-engine head-to-head vs OpenClaw, MCP search bridge into Claude/Codex/ChatGPT. The Hermes story is structurally crowding into the Loop conversation faster than any other harness.
OpenClaw — counterweight to Hermes in poker engine and most every "agent vs agent" thread this week.
Grok — model layer powering Hermes' agentic loops via Premium OAuth, no API spend.
CrewAI and LangGraph — referenced in the restaurant-analogy multi-agent post as the practical builders.
MCP (Model Context Protocol) — the recurring glue. Two of today's surviving cases use MCP to bridge agents across harnesses.
Hermes Agent (Nous Research) — central to today's signal: xAI partnership, Grok OAuth subscription path, poker-engine head-to-head vs OpenClaw, MCP search bridge into Claude/Codex/ChatGPT. The Hermes story is structurally crowding into the Loop conversation faster than any other harness.
OpenClaw — counterweight to Hermes in poker engine and most every "agent vs agent" thread this week.
Grok — model layer powering Hermes' agentic loops via Premium OAuth, no API spend.
CrewAI and LangGraph — referenced in the restaurant-analogy multi-agent post as the practical builders.
MCP (Model Context Protocol) — the recurring glue. Two of today's surviving cases use MCP to bridge agents across harnesses.
Comments