Why Enterprise AI Agents Route on Intuition Instead of Knowledge
Enterprise multi-agent systems route work by embedding similarity. Recent research shows formal ontologies outperform that approach by measurable margins.
Blog
Friday Roundup - Week 17: Postmortems, Frontier Models, and Planting Season AI
Anthropic's April 23 postmortem on three Claude Code bugs, DeepSeek V4 at 1,256 HN points, MCP architecture RCE exposure, and AI seeding prescriptions entering the 2026 planting season.
A Walkable Tour of What I Have Been Reading
Four piles of recent reading: error standards becoming agent protocols, MCP as the substrate for agentic software, data governance tooling at 1.0 and rural connectivity. They are less different than they look.
Friday Roundup - Week 16: Open Weights, Closed Sources, and Autonomous Farms
Gemma 4 achieves open-weight quality parity with closed APIs under Apache 2.0. Monarch Tractor faces shutdown reports. OpenAPI 3.2.0 gets its first complete tooling support. Claude Code ships TUI mode and push notifications.
CLI Error Messages Are a Dual-Consumer Problem in 2026
CLI tools now serve humans and AI agents from one binary. Three fixes: structured JSON output, RFC 9457 as a convention, and remediation fields.
Building a Personal Research Engine with sigint and Claude Code
How I turned a Claude Code plugin into a full research knowledge base with hierarchical organization, knowledge graphing, semantic search, GitHub Discussions, and derivative content pipelines.
Friday Roundup - Week 15: Mythos, Malware, and a Moon Return
Anthropic's Project Glasswing launches Claude Mythos to find vulnerabilities in every major OS and browser. CPUID's site served malware for six hours. The Linux kernel publishes AI contributor policy. Artemis II splashes down. Purdue's 21-year precision ag study delivers a nuanced verdict.
Friday Roundup - Week 14: Supply Chain Under Siege
Five open-source supply chain attacks in twelve days. Axios npm compromised with 100M weekly downloads. TeamPCP's lateral credential chain from Trivy to Checkmarx to LiteLLM to Telnyx to Axios. Copilot rewrote a developer's PR. The Cognitive Dark Forest as signal theory.
Context Injection: A Separate Tool for Memory-Augmented LLMs
Why memory search and context injection are different tools, and how cascading token budgets solve the prompt enrichment problem.
Friday Roundup - Week 13: Trust, Benchmarks, and the $50M Sprayer
GitHub's private repo training opt-out deadline, three supply chain attacks, Claude Code 2.1.86, ARC-AGI-3's sub-1% AI score, and Ecorobotix's $50M precision spraying investment.