Adding a Read-Only World Perception Layer to Agents
MisterMagicmike99 · reddit · 2026-07-16
The author developed a world-sense MCP for agents, allowing models to read the current world state via 10 read-only tools for reasoning and comprehensive judgment.
Tool data is sourced from a private world-engine (A.R.C.A.N.E.), covering 6 risk domains: finance, energy, society, military, cyber, and weather. Each domain provides not only current risks but also change trajectories over the last 90 days, along with read interfaces for "what changed today vs. yesterday/last week/last month."
The author emphasizes that it is neither a predictor nor an action-taker:
- Read-only design; tools only return pre-published values.
- Executes no code and does not actively scrape external content.
- Sanitizes third-party texts to prevent potential prompt injection.
- Employs allow-list projection, fail-closed leak protection, a kill switch, and rate limits.
Two use cases are provided:
- Letting a personal agent in a daily cron check risks and health alerts based on travel itineraries.
- Serving as structured world-info input for analytical agents to reduce hallucinations caused by scraping the web.
The author also published the request path code and safety guidelines, encouraging public auditing.
More from coding & agent
- An MCP server signs every AI agent tool call into a verifiable Merkle chain — Funky_Chicken_22 · 2026-07-22
- Claude Code skill uses 10 Markdown rules to make outputs ADHD-friendly — alex_verem · 2026-07-22
- A Firecracker-based platform says it can host 6,000 AI agents on one 256 GB server — maritime_sh · 2026-07-22
- A better path to agent autonomy is running waves, finding friction, and iterating — JnBrymn · 2026-07-22
- AI agent designers map the visual and tonal cues behind companionship products — Unlikely-Platform-47 · 2026-07-22
- Coding agents are heading toward an AI-writes, AI-reviews, human-approves workflow — aftahi_ai · 2026-07-22