"Nous Research Blank Slate: build agents with zero default tools"
Why this is in the vault
Hermes Blank Slate's opt-in capability model is a direct architectural counterpart to the RDCO harness design principle of minimizing default surface area — and a signal that zero-trust tool provisioning is becoming a first-class agent design pattern.
Issue contents
Top News — Third-party
Nous Research shipped a "Blank Slate" setup mode for Hermes Agent. The agent starts with only four defaults: a provider, a model, file operations, and a terminal. Web access, browser control, code execution, vision, memory, and MCP servers are all disabled by default and remain off across updates. Capabilities are added back explicitly via hermes tools or hermes skills commands. The design inverts the current-tool-default norm — nothing is on until you ask for it.
Top Repo — Third-party Nous Research also released Hermes Agent v0.17.0 with significant reach expansion: iMessage via Photon Spectrum (OAuth, no Mac relay required), background subagents (fire-and-keep-chatting pattern), Automation Blueprints (plain-English task scheduling), xAI Grok Cursor Composer model, WhatsApp Business Cloud API, and Raft agent network support. 1,475 commits, 245 contributors at release.
Top Paper — Third-party UC Berkeley released LOCUS, a dataset of 2.2M U.S. local laws scraped from 9,239 cities and counties. Each law is scored on paternalism, opacity, enforcement discretion, and salience. Available on Hugging Face as LocalLaws/LOCUS-v1. OCR cost cited at $0.30/1,000 pages.
Signals (brief items) — Third-party
- New tool synthesizes non-natural "alien" programming languages that LLMs reportedly execute more accurately than English instructions
- Stanford releases autonomous research agent that writes cited reports (29K GitHub stars)
- Five major LLMs show pro-female hiring bias across cultural contexts including Japanese resume sets
- Critic-free training method beats GRPO on hard math benchmarks by 18%
- Curated list of open-source replacements for paid developer tools
No self-cross-promo detected. House ad is a standard "promote to 250K+ AI developers" boilerplate — not a named sponsor.
Mapping against Ray Data Co
The Blank Slate pattern is immediately relevant to the RDCO harness posture. The RDCO Claude Code setup adds MCP servers and skills incrementally; Hermes Blank Slate formalizes that discipline at the agent runtime layer — nothing is on until explicitly provisioned. This is the same "thin harness, fat skills" principle applied to tool surface area.
The v0.17.0 background subagents feature (fire-and-keep-chatting) maps to the RDCO multi-agent fan-out pattern used in /deep-research and /process-newsletter. Worth watching whether Hermes provides session isolation guarantees comparable to worktree isolation.
The iMessage OAuth integration (no Mac relay) is directly interesting given the channels-agent Mac Mini setup — Photon Spectrum may be a lighter-weight path than the current always-on tmux relay approach.
LOCUS (local law dataset) is not directly relevant to current RDCO bets (phData DSA, Markov capital-cycle) but could surface for a client engagement in compliance or regulatory data engineering.
No items in this issue touch prediction markets, automated investing, or content-as-product.
AlphaSignal is on the F (follow-forward only, no backfill) list — file and move on.
Related
- [[2026-05-21-mcp-vs-cli-agent-native-interface-task-class]] — task-class split for agent-native interfaces; Blank Slate's explicit provisioning model maps to the CLI-side of that split
- [[2026-04-07-claude-code-architecture-teardown]] — harness architecture teardown; the "remove harness complexity" design principle aligns with Blank Slate's zero-default philosophy
- [[research/2026-06-04-a2a-protocols-beyond-mcp-rdco]] — A2A vs MCP for inter-skill handoff; Hermes v0.17.0 Raft agent network is a data point here
- [[2026-04-28-alphasignal-claude-creative-mcps-platonic-hypothesis]] — prior AlphaSignal issue covering MCP creative tooling