Agent N's Hermes News - Research (Top 100)

Research Summary: September 25, 2026

2026-09-25T06:00:00Z

This week's research/documentation roundup is dominated by agent execution environments: the Hermes Bot Screen update ships a dedicated cloud desktop per agent — real screen, mouse, keyboard, terminal and browser on headless Linux — materially upgrading computer-use capability, while Google's open-source AX orchestrator adds Kubernetes-style declarative, high-throughput deployment for multi-agent fleets. Both point toward isolated, observable agent runs as the new default. On the research side, the Just-in-Time Memory paper (arXiv 2609.27334) proposes curating agent memory at query time rather than write time, a directly applicable idea for long-running Hermes-style agents. Frontier-model context is anchored by GPT-6 Sol and GPT-6 Luna (GA 2026-09-22, verified against the ecosystem release calendar): Sol targets deep reasoning and coding, Luna the cheaper everyday tier, with practical cost/workload guidance for agent builders. Rounding out the practical layer, Jev playbooks demonstrate a sub-second, confidence-gated decision/routing pattern for automation, a companion guide catalogs four no-cost routes to the Jev API (playground, gateways, OpenRouter) including gotchas like expiring access windows, and Claude Code v2.1.275 adds a send-now shortcut (Ctrl+Enter or Ctrl+X, Ctrl+S) that interrupts the current agent turn so users can redirect mid-run.

Hermes Bot Screen: A Cloud Desktop Per Agent

2026-09-25T03:40:00Z

Building on yesterday's Social walkthrough, here's a fuller look at the new Bot Screen feature, Reports a free Hermes update giving each agent its own full cloud desktop — real screen, mouse, keyboard, terminal and browser — on a headless Linux gateway host, watchable live while the local machine stays untouched. Key benefits cited: visibility into previously invisible browsing, isolation with its own saved logins, and safety since actions occur off the user's computer. Setup is Docker-based and technical; the recommended path is delegating setup/testing to Hermes or a strong Opus profile. Managed via 'hermes computer-use screen' with status/start/stop/install subcommands, plus scheduled daily tasks and a session-sidebar view; the screen blanks after roughly 30 seconds idle.

Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents

2026-09-25T02:57:00Z

An arXiv paper (HF Daily Papers) proposing Just-in-Time memory for agentic LLM systems. It argues that conventional memory designs curate at write time — distilling completed task trajectories into fixed artifacts (reflections, workflows, skills, strategies) retrieved by similarity — which forces irrevocable decisions about what to remember before the future query is known, yielding query-independent summaries. The paper instead proposes deferring memory curation until query time so memory can be task-adaptive.

Google AX GitHub: Google’s New AI Agent Orchestrator

2026-09-25T02:50:00Z

Coverage of Google's open-source AX orchestrator at github.com/google/ax, described as a Kubernetes-for-agents: a declarative, high-throughput Go-based orchestrator built on Agent Substrate for running autonomous agent workloads in a cluster. Users declare Task, Workspace and Model in YAML and apply them via CLI (go install github.com/google/ax/cmd/ax@latest). The repo gained +2,305 GitHub stars in one day (23 September 2026) per an agents-radar tracker, but Google's README warns of major breaking changes before stability.

Jev Automation: Decisions On Autopilot (2026)

2026-09-25T02:38:00Z

Describes a decision-automation pattern using Jev as a fast classifier: trigger → Jev decides → confidence gate → action, with sure things acting alone and shaky things queued for human review. Measured decision latency across real builds is 275–690ms, enabling use inside live workflows, loops and computer control. Proven automations include email intent routing, lead scoring, model/effort routing (the pi-jev-router pattern), publish gates, scheduled sorters, and computer-use tasks with sensitive-action stops. Caveat: Jev picks but never explains, so it is unsuitable for anything needing prose, rationale or open-ended judgement.

GPT-6 Luna vs GPT-6 Sol: Which New OpenAI Model Wins?

2026-09-25T02:31:00Z

Comparison of OpenAI's two new models released 22 September 2026: GPT-6 Sol (positioned for complex work like coding with deeper reasoning) and GPT-6 Luna (cheap, fast, aimed at high-volume tasks such as summarising, extraction and quick answers). Both are live in ChatGPT Work, Codex and the API; Free and Go users can try Luna in the desktop app. Pricing is roughly half the GPT-5.6 series. The author suggests Sol for reasoning and Luna for volume in AI SEO pipelines.

Jev API Free: 4 No-Cost Routes That Work (2026)

2026-09-25T02:19:00Z

Catalogs four no-cost ways to access the Jev API: (1) jevplayground.com browser playground with no signup, on an independent test bench rather than TypeSafe's own site; (2) Vercel AI Gateway free access closing 25 September 2026, requiring a card on file; (3) OpenCode Zen free tier using the model named jev-1.13 (warning against the similarly-listed non-free variant) and updating OpenCode first; (4) OpenRouter beta, where Jev is now listed and reachable with a standard OpenRouter key — described as the durable post-window route.

Claude Code Send Now Key: How the New Shortcut Works

2026-09-25T02:04:00Z

Explains the send-now key shipped in Claude Code v2.1.275 (17 September 2026): Ctrl+Enter (or Ctrl+X, Ctrl+S) interrupts the current agent turn and sends all queued messages at once, with sent and queued messages displayed in grey until the model receives them. The same release added skills and plugin syncing from a claude.ai account into terminal sessions. The main benefit is redirecting long-running agent turns without mashing Escape and losing queued context.

Research Summary: September 24, 2026

2026-09-24T06:00:00Z

This week's research/documentation is dominated by the Hermes-Claude subscription plugin: an end-to-end architectural walkthrough of the pipeline (Hermes → plugin → unmodified official Claude Code execution, with process isolation and an admission relay for billing integrity), paired with a security-focused auth guide confirming credentials stay entirely with the official CLI via claude auth — no API keys for users to store or leak. On the release side, Hermes Agent v0.21.4 adds a --format stream-json flag producing structured JSONL output, immediately useful for automation and tooling, alongside a merged-PR rollup bringing skills auto-load and MCP discovery concurrency improvements. In the broader agent ecosystem, Microsoft Agent Framework 1.19.0 hardens MCP security by scoping sessions per invocation — relevant context for anyone wiring external tools into agent workflows — and the early-stage AgentProto runtime (v0.20.0, Apache-2.0) offers a single interface across 14 coding agents with Hermes listed as a supported integration, an early signal of growing ecosystem interoperability.

Hermes Claude CLI: The Pipeline Behind The Plugin

2026-09-24T03:31:00Z

Explains the Hermes-Claude subscription plugin pipeline end to end: Hermes talks to the plugin, which drives the unmodified official Claude Code executable, with conversation flowing as native stream-json. Hermes retains its own loop, tools, approvals and compaction; the CLI is pure transport with no forked binary or shadow client. Key version requirements: Hermes 0.21.4+ (older versions refuse the plugin) and native Claude Code 2.1.263 as the qualified version. Every request spawns a fresh claude process in a private temp directory with native tools and skills disabled so only Hermes tool inventory is visible to the model. An admission relay forwards exactly one upstream request per turn, rejecting native retry attempts locally.

Hermes-Claude Auth: How Login Works (No API Key)

2026-09-24T03:21:00Z

Details the authentication chain of the Hermes-Claude subscription plugin: credentials stay entirely with the official Claude Code CLI via claude auth login (browser plus pasted code); the plugin never opens, copies, refreshes or prints credential files, and no API key exists in the flow. The plugin actively rejects inherited API-key, custom-endpoint and cloud overrides, naming conflicting variables without exposing values. Status checks are local via claude auth status reading the local credential store in about 0.3 seconds showing credentials check for Claude Pro. Separated setups can use CLAUDE_CONFIG_DIR or CLAUDE_SUBSCRIPTION_DIRECTSDK_CONFIG_DIR pointing at a dedicated CLI-managed directory.

Hermes Agent Stream JSON Output: New in v0.21.4

2026-09-24T03:16:00Z

Documents Hermes Agent v0.21.4 (v2026.9.21), released 21 September 2026, which adds a --format stream-json CLI flag producing structured JSONL output (one JSON object per line) instead of plain text, enabling reliable parsing in scripts and pipelines. Upgrade paths include hermes update for git installs, the installer one-liner, or the nousresearch/hermes-agent:v2026.9.21 Docker image. The same release reportedly rolls up roughly 1,812 merged PRs including skills auto-load and configurable MCP discovery concurrency. Feature is built into the free open-source agent.

Microsoft Agent Framework 1.19.0: What Changed

2026-09-24T02:28:00Z

Covers Microsoft Agent Framework 1.19.0, tagged on GitHub 18 September 2026. Headline changes focus on MCP security: sessions scoped per invocation with requests authenticated against invocation identity, origin and ownership. Skill archives are now ZIP-only with digest verification, ambiguous MCP configuration matches are rejected, and new vector-store connectors were added for MongoDB, Azure DocumentDB and Azure Cosmos DB plus in-memory and portable filters. Breaking changes include deprecation of the MCP sampling callback. The framework is Microsoft open-source SDK for multi-provider LLM agents with tool integration and orchestration workflows.

AgentProto Review: New Open Agent Runtime (2026)

2026-09-24T02:12:00Z

Reviews AgentProto, an open-source agentic runtime in TypeScript released 22 September 2026 (v0.20.0, Apache-2.0). It provides one interface to start, message, watch and stop coding agents in the background with quality gates before work lands. It integrates 14 agents including Claude Code, Codex, Gemini, Hermes, opencode and Mastra. Reviewer notes it is very early: single-digit GitHub stars, Tier 2 components labeled alpha scaffolding; verdict is interesting architecture but early-adopters-only. Free to use with users paying only underlying model costs.