swarm-master/swarm-common/obsidian-vault/will/will-shared-zap/Atlas/Daily Research/2026-05-20 - Hermes AI Brief.md

Daily Hermes + AI Research Brief — 2026-05-20

Important updates

• Hermes Agent v0.13 / “Tenacity” remains the key Hermes update to track. GitHub release/search results list v2026.5.7 with Kanban as a durable multi-agent board, heartbeat/reclaim/zombie detection/auto-block behavior, and a large reliability-focused release since v0.12.0. This matters because Will’s Atlas setup already depends on default gateway stability plus specialist profiles/Kanban for durable work. Source: NousResearch/hermes-agent releases.
• Hermes docs continue to emphasize self-improving skills, persistent memory, profiles, gateway, plugins, MCP, cron, and Kanban. For Will, the practical takeaway is to keep Atlas’ daily operations split between short memory pointers, Obsidian for durable context, and specialist-profile delegation rather than stuffing everything into one long session. Source: Hermes Agent Documentation.
• OpenAI Codex CLI shipped another May update. The Codex changelog shows “Codex CLI 0.130.0” on 2026-05-08 with performance improvements/bug fixes. This is relevant because Will’s Hermes workers/profiles use Codex auth in some paths; keep profile auth smoke tests in the loop before dispatching autonomous coding agents. Source: OpenAI Codex changelog.
• 1Password announced/covered just-in-time credential access for Codex via MCP. The useful pattern is not the vendor hype; it is the architecture: coding agents should request narrowly scoped, auditable secrets at task time instead of having broad static env access. This maps directly to safer Hermes MCP/tool credentials and swarm service secrets. Source: SiliconANGLE coverage.
• MCP remains the practical integration layer to watch. Recent MCP comparisons and Anthropic engineering material emphasize implementing tools once and exposing them across agent clients. For Will, this supports consolidating local services—Obsidian, n8n, Ollama/llama.cpp, Kokoro/Whisper—behind stable MCP/tool wrappers instead of one-off scripts. Sources: ClickHouse MCP framework comparison, Anthropic on MCP code execution.
• AI observability is moving from infra-only metrics to LLM-specific monitoring. Current LLMOps guidance stresses that healthy CPU/GPU dashboards do not prove agent quality; you need latency, error rates, tool-call failures, prompt/model routing traces, and evaluation signals. This is directly useful for Will’s CoreWeave/GPU/k8s-style work and local swarm reliability. Source: Kong AI observability guide.

Actionable ideas for us

• [quick] Add a recurring manual check item for Hermes release notes: compare local ~/.hermes/hermes-agent against origin/main and v2026.5.7+ release notes, but only use the safe isolated update flow.
• [quick] Smoke-test specialist profiles that may use Codex: hermes -p <profile> chat -q 'Reply exactly: ok' --toolsets safe -Q before Kanban dispatch.
• [experiment] Prototype an MCP-style “just-in-time secret” pattern locally: agent requests a named credential lease for one task; logs scope and expiration; never exposes raw secrets in final output.
• [watch] Track Hermes Kanban reliability and session-rollover/handoff changes; these are likely to matter more for Atlas autonomy than flashy model announcements.

Worth ignoring

• Generic “AI agents in 2026” listicles with no implementation detail.
• Funding/partnership headlines unless they ship concrete APIs, MCP servers, model routing, or local-first tooling.
• Consumer chatbot feature news with no path to Hermes, Obsidian, n8n, local inference, or GPU ops.