Saturday, November 7, 2026 270 picks · 133 builders credited Subscribe
AgentRags.

The agentic rag, wrung out.

The rag for

Tuesday, October 6, 2026

The day in one breath

Hermes Index. Nous Research put a public scoreboard on model performance inside Hermes Agent, with cost per task next to the score. Opus 5.5 leads the official card; Astra is second and much more expensive per task; Sonnet 5.5 is the mid-tier efficiency pick on the numbers they posted. It is the story builders will argue about, and the one Fake Peter should lead with.

Industry

Hermes Index: same harness, score and dollars per task

Nous Research launched Hermes Index, a mean of four suites run inside the same Hermes Agent harness with reasoning set high where the model offers it. Suites: new Hermes Bench (150 tasks, 25 categories: skills, research, diagrams and creation, memory, tool use, safety), Terminal-Bench 4.0, Terminal-Bench Science, and SkillsBench. Graders look at workspace files, task state, and tool use. Portal: https://portal.nousresearch.com/bench. Official first board: Claude Opus 5.5 at 63.31 and $4.99/task, GPT 6 Astra at 56.25 and $11.61, Sonnet 5.5 at 53.14 and $2.82. Cheap end from the same thread: DeepSeek V4.1 Flash 36.91 at $0.26, Ling 3.0 Flash 21.56 at $0.054. Community recaps also put GPT 6 Sol near 44.10 / $2.23 and Grok 4.7 near 39.32 / $10.77, and flag some Opus and Sonnet cells as provisional. Treat those extras as secondary until checked on the portal.

Credit@NousResearch

Guides & Workflows

Ask Hermes to find and install a skill

Hermes Wingtips #93: the agent can search for a skill and run the install itself. Example prompt: “find a skill for flashcards and install it.” It picks a match, runs the security scan, then tells you what to say to use the skill. Aimed at fresh profiles and at people who do not know the skill name or the install command.

Credit@witcheer

Releases

142 merges on 5 Oct: usage-wall chip, plugin slash skills, dashboard login hardening

Luke The Dev’s digest of 142 PRs merged 5 Oct. Operator set: Desktop shows a progress chip before a subscription hits its usage wall, and rate-limited providers say why and until when in model pickers. Settings gets one Plugins pane. Plugin skills appear as `/plugin:skill` in TUI, Desktop, and CLI and can be stacked. `hermes chat -q` no longer hangs on approval, memory, MCP, or vault prompts. `hermes update` no longer stalls on a 20-minute git repack. Gateway reloads a platform whose plugin failed at startup without a full restart; Telegram flushes messages held during a disconnect. Dashboard login caps request bodies, bounds login values, and ignores spoofed `X-Forwarded-For`. Catalog memory providers (hindsight, mnemosyne) stop failing on every start, and installing one actually activates it. Memory writes on hand-formatted files fail with a fix hint instead of a silent miss. Opus 5.5 keeps thinking on; Anthropic summarized thinking is no longer promoted to the final answer. New hooks for human-input requests, post-gateway admission, and which cron run you are in.

Credit@iamlukethedev, @HermesWatcher

Security

Exposed Docker plus a planted Hermes Agent, API keys first

Recap of ThreatDown’s CARBONATO report (dated in the post as 22 Sep). Claim: bots hit unauthenticated Docker API on port 2375, drop a stock Hermes Agent, and replace the instruction file with a 39-line brief that harvests AI API keys before SSH keys or databases. Named targets include OpenAI, Anthropic, Google, Ollama, and vLLM. Stolen keys were pointed at an attacker-run API that advertised 12 models and served 27. Advice in the post: do not publish Docker’s API without auth; rotate AI keys like bank credentials. Primary write-up is the vendor report, not this post.

Credit@Rakhsh_Tech

Community

User-story page hits 371, with a work batch

Update only. The docs page announced yesterday (326 stories) now has 45 more, mostly work use. Examples called out: Razorpay giving 220+ employees their own always-on agent; a pool company flagging bad technician reports while the route is live; a physician saving 30–45 minutes of documentation per case; a Swiss agency drafting invoices a human still sends; payment reconciliation for a local business; a bookstore running one agent for staff and one for customers. Page: https://hermes-agent.nousresearch.com/docs/user-stories. Nous reposted the thread early 7 Oct UTC.

Credit@witcheer

Never miss a night

Every pick, delivered to your reader.

One feed for everything, or one per agent harness you run. Rebuilt every night, with full credit on every item.