Hermes Ships Judgment Release; MCP Goes Stateless Before July 28

<p><strong>Hermes Agent v0.18.0</strong> shipped July 1 with zero open critical bugs, Mixture-of-Agents as a first-class virtual model, and a <code>/learn</code> command for on-the-fly skill distillation — already the top app on OpenRouter by token volume. The <strong>MCP 2026-07-28 specification RC</strong> is now active, making the protocol stateless at...

Highlights

  • Hermes Agent v0.18.0 — “The Judgment Release” — shipped July 1 with zero open P0/P1 bugs, Mixture-of-Agents as a first-class virtual model, and a new /learn command for on-the-fly skill distillation (Nous Research)
  • The MCP 2026-07-28 specification release candidate is active and finalizes in four weeks, making the protocol stateless at its core and introducing six breaking changes for production server operators (MCP Blog)
  • DeepSeek V4-Pro scores 80.6% on SWE-bench Verified under an MIT license — the highest published open-weight result for agentic coding, within 0.2 points of top closed models (DeepSeek Guide)
  • Anthropic’s Zero Trust for AI Agents framework is gaining enterprise traction in July, establishing cryptographic agent identity and least-privilege tool grants as the new production baseline (Anthropic)

Key Signals

  1. July 1, 2026

    Hermes Agent v0.18.0: zero bugs, Mixture-of-Agents, self-improving skills

    Nous Research closed ~1,950 issues and PRs in this release — every P0 and P1 cleared — and promoted Mixture-of-Agents to a first-class virtual model: teams can now slot a named ensemble like “my-council” into any model picker alongside Claude, GPT, or Grok, with each constituent model’s full output rendered before synthesis. The /learn command lets the running agent distill a reusable skill from a directory, URL, or a live workflow it just watched you complete. Hermes is already the top OpenRouter application by token volume at 17 trillion tokens processed.

    (GitHub Release, OpenRouter)

  2. RC active; final Jul 28, 2026

    MCP 2026 specification RC: stateless core, four weeks to final

    The MCP team locked the 2026-07-28 release candidate on May 21. The headline change: the protocol is now stateless at the transport layer — a remote MCP server that previously required sticky sessions, a shared session store, and deep-packet inspection at the gateway can now run behind a plain round-robin load balancer, routing on an Mcp-Method header with client-cacheable tool lists. Six Specification Enhancement Proposals (SEPs) land together: MCP Apps (server-rendered sandboxed HTML UIs), a Tasks extension for long-running operations, and OAuth/OIDC-aligned authorization. Production server teams have four weeks to audit and adapt before the final spec drops.

    (MCP Blog, ChatForest breakdown)

  3. Released Apr 24, 2026

    DeepSeek V4-Pro: MIT-licensed frontier coding at 80.6% SWE-bench

    V4-Pro is a 1.6T-parameter MoE model (49B active) with a 1M-token context window, scoring 80.6% on SWE-bench Verified in Think Max mode — the highest published result for any open-weight model and within 0.2 points of Claude Opus 4.6. The MIT license means commercial hosting, fine-tuning, and redistribution with no restrictions. For teams running coding or agentic pipelines at scale, this meaningfully reopens the closed-vs-open cost trade-off.

    (DeepSeek Guide)

Why It Matters / What To Watch

  1. Open-source agent and model stacks are closing the gap with managed platforms

    • Hermes v0.18.0 ships persistent memory, 40+ built-in tools, Mixture-of-Agents ensembles, a desktop app with git worktree management, and a scheduling gateway — capabilities previously reserved for managed commercial platforms. Teams evaluating Bedrock AgentCore or similar offerings should benchmark against the OSS tier now. (GitHub Release)
    • DeepSeek V4-Pro under MIT means frontier agentic coding is now fine-tunable and self-hostable. If your pipeline’s primary cost is closed-weight inference for code tasks, the calculus deserves a fresh look. (DeepSeek Guide)

  2. MCP production operators have a hard deadline and breaking changes to ship

    • The six SEPs in the MCP 2026-07-28 RC require externalizing session state from servers that previously held it in-process — the most disruptive change for existing deployments. Tier 1 SDK maintainers are expected to ship support within the ten-week validation window ending July 28. (MCP Blog, ChatForest)
    • The Anthropic Zero Trust framework published May 27 is increasingly referenced as the enterprise checklist for agent identity and tool scoping. As the MCP ecosystem grows past 500 public servers, runtime authorization is moving from optional to expected. (Anthropic, Adversa AI)

Quick Links