swarm-master/swarm-common/obsidian-vault/will/will-shared-zap/Atlas/Daily Research/2026-05-19 - Hermes AI Brief.md

Daily Hermes + AI Research Brief — 2026-05-19

Important updates

• Hermes Agent v0.14.0 / v2026.5.16 is the main Hermes item to track. GitHub release search shows a May 16 release with 808 commits, 633 merged PRs, and 1,393 files changed since v0.13.0; snippets highlight PyPI install support, ~19s faster cold start, much faster Browser CDP calls, and new messaging work. This matters because Will’s default Atlas gateway is source-installed and gateway uptime/update safety is more important than chasing the release immediately. Source: GitHub releases / release search result.
• Hermes v0.13.0 / v2026.5.7 shipped the “Tenacity” durable-work direction. NewReleases summarizes it as Kanban becoming a durable multi-agent board with heartbeat, reclaim, zombie detection, and auto-block behavior. This aligns directly with Will’s specialist-profile/worker setup; it is worth treating Kanban as the safe path for long agent tasks instead of ad-hoc background spawns. Source: NewReleases v2026.5.7.
• Hermes docs continue to emphasize self-improving skills, persistent memory, profiles, messaging gateway, and provider-agnostic routing. That is not a “new today” item, but it confirms Atlas’s current architecture choices: keep default as the production Telegram gateway, use specialist profiles for isolation, and save durable research outputs into Obsidian rather than memory. Source: Hermes Agent docs.
• MCP scaling pattern worth adopting: expose many tools as code APIs inside execution environments, not as hundreds of direct LLM tools. Anthropic’s engineering writeup says code execution with MCP can make agents more efficient by letting the model write code that calls MCP APIs, reducing tool-schema/token pressure. For Will, this suggests future Atlas/n8n/local-swarm integrations should prefer compact wrapper APIs and executable client libraries when tool count grows. Source: Anthropic Engineering — Code execution with MCP.
• Open-source infra maintainers are now explicitly dealing with AI-generated PR load. MLSys has an invited talk, “Rethinking Open Source Contribution in the Age of AI Agents,” framed around vLLM and the surge of AI-generated pull requests. This matters for Will’s CoreWeave/k8s/LLM-infra work: review gates, narrow tests, provenance, and anti-slop contribution policies are now part of production LLM ops, not just repo hygiene. Source: MLSys 2026 schedule / invited talk.

Actionable ideas for us

• [quick] Check local Hermes safely before updating: hermes --version, git status --short --branch, and git rev-list --left-right --count main...origin/main; do not run hermes update automatically if the tree is dirty.
• [quick] Review whether Atlas’s Browser/CDP and messaging paths benefit from v0.14.0, but route any upgrade through the safe update workflow because the gateway is production.
• [experiment] Prototype one “code API over tool flood” integration for local swarm services: a small Python client that wraps n8n, llama.cpp, Ollama embeddings, and Obsidian REST behind a few stable calls.
• [watch] Track Hermes issues/releases around v0.14.0 for Windows/PyPI/lazy dependency fallout and any gateway regressions before adopting it on default.

Worth ignoring

• Generic “agent landscape 2026” listicles unless they include concrete implementation details, benchmarks, or repo links.
• Funding/market stories about AI agents with no deployable tooling, protocol, model, or infra takeaway.
• Consumer-only agent announcements unless they expose useful MCP/tooling/local-first patterns.