swarm-master/swarm-common/obsidian-vault/will/will-shared-zap/Atlas/Daily Research/2026-05-30 - Hermes AI Brief.md

Daily Hermes + AI Research Brief — 2026-05-30

Important updates

• Hermes shipped v0.15.2/v0.15.1 hotfixes on May 29. v0.15.2 fixes packaging so bundled plugin.yaml manifests ship in wheels/sdists; v0.15.1 fixes the v0.15.0 dashboard infinite-reload loop in loopback/Docker/hosted setups, restores .md media delivery, fixes kanban worker SIGTERM, Docker MCP npx/npm/node PATH resolution, /yolo session bypass, and expands the skills catalog from 858 to 19,932 entries. This is directly relevant to Atlas gateway/dashboard reliability and Dockerized MCP servers. Source: Hermes releases.
• Hermes v0.15.0 “Velocity Release” is a major architecture + swarm release. The core run_agent.py was split from ~16k LOC to ~3.8k across agent/*; kanban gained auto-decomposition, swarm topology, worktree-per-task, per-task model overrides, scheduled starts, TTL/retry/stale-task handling, and inspection endpoints; session_search was rebuilt as no-LLM/free/FTS-based and reported ~4,500× faster. This is worth folding into Atlas workflows instead of bespoke swarm glue where possible. Source: Hermes v0.15.0 release notes.
• Hermes added promptware / Brainworm-class defenses. v0.15.0 release notes describe scanning recalled memory, tool output delimiter hardening, and centralized threat patterns. For Will’s local-agent work, this is a good reason to route more risky web/file recall through Hermes-native surfaces rather than ad hoc prompts. Source: Hermes releases.
• Kubernetes LLM serving is converging on inference-aware routing. Datadog’s May 29 writeup explains Kubernetes Gateway API Inference Extension routing based on backend state such as KV-cache readiness, LoRA adapter availability, queue length, health/readiness, and body-based model routing. This maps well to CoreWeave-style GPU/k8s work and any future local swarm router that should avoid naive round-robin. Source: Datadog: Monitor LLM routing with the Kubernetes Inference Extension.
• Anthropic continues pushing MCP into first-party agent APIs. Search results for Anthropic’s “New capabilities for building agents on the Anthropic API” highlight an API-side MCP connector so developers can connect Claude to remote MCP servers without writing custom MCP clients. This reinforces MCP as the default tool boundary to support in Atlas/Hermes integrations. Source: Anthropic agent API capabilities.
• Agent eval loops are becoming a practical norm. OpenAI’s developer materials now frame agent improvement around traces → human/model feedback → evals → harness changes, where the “harness” includes instructions, tools, routing, output requirements, and validation checks. That is a useful template for Atlas regression tests and cron-job quality checks. Source: OpenAI cookbook: Agent improvement loop.

Actionable ideas for us

• [quick] Upgrade/check Hermes to at least v0.15.2 before doing dashboard, Docker, MCP, or kanban work; the May 29 hotfixes address exactly those surfaces.
• [experiment] Replace one Atlas swarm prototype with Hermes kanban swarm primitives: per-task worktrees, model overrides, TTL/retry, and verifier/synthesizer gates are now built-in enough to test against Will’s existing Pi/local-agent harness.
• [experiment] Add a small inference-routing design note for local/k8s agents: track queue length, model/adapter residency, KV-prefix reuse potential, and readiness; compare naive routing vs inference-aware routing.
• [watch] Build lightweight eval traces for this daily brief job: store source queries, chosen links, rejected hype, and final bullets so future Atlas can measure “useful to Will” rather than just successful delivery.

Worth ignoring

• Generic “top AI developer tools in 2026” listicles unless they include concrete tool APIs, eval methodology, or deployment patterns.
• Broad MCP-is-dead / MCP-is-everything takes; the actionable signal is implementation quality: auth, server discovery, sandboxing, and observability.
• Consumer/enterprise-positioning announcements without reproducible technical details or local-agent relevance.