Recursive Self-Improvement: The Latest From Anthropic
Anthropic’s recent essay frames recursive self‑improvement (RSI) as a concrete engineering loop where AI designs, tests, and deploys its successors. The company shows that Claude‑generated code grew from a few percent to over 80% of merged changes by May 2026, and engineers now merge roughly eight times more code per day than in 2024. While the coding stage of the loop is maturing, Anthropic warns that evaluation, safety, and human oversight remain critical bottlenecks. The piece argues that each automation layer shifts the limiting factor from typing to review, judgment, and governance.
Microsoft's Approach to LLM: MAI-Thinking-1
Microsoft’s AI team released a technical report on MAI‑Thinking‑1, a reasoning model that achieved 52.8% on SWE‑Bench Pro and 97% on AIME 2025, matching frontier‑size models. The model was trained on 30 trillion human‑written tokens without any external model distillation, emphasizing learning...
CLAUDE CODE ORCHESTRATION
Claude Code’s new orchestration model introduces three primitives—Dynamic Workflows, Subagents, and Agent Teams—to let developers match execution style to problem complexity. Dynamic Workflows auto‑generate JavaScript scripts that can spawn up to 1,000 parallel agents, run adversarial verification, and return only...
Self-Evolving Agent Skills: SkillOpt
SkillOpt, a new framework from Microsoft and leading Chinese universities, treats agent instructions as trainable external state, applying deep‑learning‑style optimization to text. By imposing a bounded edit budget—called a textual learning rate—the system performs controlled “textual gradient descent,” preventing erratic...
An Implementation Checklist to Claude Code in Large Codebases
The post translates Anthropic’s Claude Code best‑practice guide into a step‑by‑step checklist for large codebases. It emphasizes building a harness—root and subdirectory CLAUDE.md files, targeted hooks, and on‑demand skills—to give Claude the right context without maintaining an index. Each phase...
GCIS 2026: My Agentic-AI Lens on a Prestigious, Invitation-Only Cyber Summit Near Washington
The Global Cyber Innovation Summit (GCIS) in February 2026 gathered senior cyber and AI leaders to discuss the emerging agentic‑AI era. Attendees warned that trillions of autonomous agents will soon overwhelm existing infrastructure, while weaponized models can turn a disclosed...
Claude Agents Can Now Dream: How AI Engineers Should Use Anthropic’s New Agent Features Without Creating New Attack Paths
Anthropic unveiled three new capabilities for Claude Managed Agents—Dreaming, Outcomes, and Multi‑agent Orchestration—shifting the platform from single‑threaded assistants to a managed, inspectable agent operating system. Dreaming lets agents synthesize lessons from up to 100 past sessions into a separate memory...
What a Secure Harness for Agentic AI Actually Is
Enterprise teams are conflating terms like guardrails, gateways, and governance, leaving a critical gap in securing autonomous AI agents. A "secure harness" is defined as an engineered control layer that provides visibility, policy enforcement, and real‑time intervention across an agent’s...
Why Your Agentic AI Pentester Is Probably Just a Fancy Scanner
Ridge Security benchmarked three agentic AI pentesting platforms—RidgeGen, Shannon and Strix—against the OWASP Juice Shop. Using the same Gemini 3 Flash model, RidgeGen produced 55 fully verified findings with zero hallucinations, while Shannon generated 27 findings of which 63% were...
World Models, Architectures, and the Next Phase of AI
At the Spring School AI For Impact, Yann LeCun and Eric Xing sparred over the core design of world models: LeCun’s Joint Embedding Predictive Architecture (JEPA) predicts future latent representations without reconstruction, while Xing’s Generative Latent Prediction (GLP) re‑attaches a decoder to...
The Computational Wall: Why the Defense Trilemma and the NP-Hardness of Reward Hacking Detection Demand a New Security Posture for...
At a National Academies panel, researchers presented two converging impossibility results: the Defense Trilemma shows that wrapper defenses around LLMs cannot simultaneously guarantee continuity, utility preservation, and complete safety, and recent proofs demonstrate that detecting reward‑hacking is NP‑hard. Both findings...
Chapter 15: Structured Output and Schema-Constrained Generation (Claude Code Vs. Hermes Agent)
The post explains structured output—a technique that forces LLMs to return validated JSON instead of free text—and compares two implementations: Claude Code’s `jsonSchema` parameter and Hermes Agent’s tool‑forcing approach. Claude Code wraps the schema as a synthetic tool, tracks retries...
Chapter 14: Model Routing and Provider Abstraction (Claude Code Vs. Hermes Agent)
Model routing bridges the gap between an agent’s need for a language model and the actual API call, handling provider choice, format translation, context limits, and fallback. Claude Code implements this as a compile‑time TypeScript layer with static rules and...
Chapter 13: MCP Integration — Connecting Agents to the World (Claude Code Vs. Hermes Agent)
The Model Context Protocol (MCP) is emerging as a universal standard that lets AI agents invoke external tools without custom code. Claude Code embeds an MCP client directly into its TypeScript QueryEngine, using double‑underscore namespacing and dynamic discovery. Hermes solves...
Chapter 12: The Skill System Pattern (Claude Code Vs. Hermes Agent)
The chapter introduces the Skill System Pattern, contrasting Claude Code’s minimal CLAUDE.md approach with Hermes Agent’s full‑featured skill subsystem. Claude Code simply discovers static markdown files at session start, lacking versioning, creation, security scanning, or a marketplace. Hermes, by contrast,...