Introducing STATE-Bench: A benchmark for AI agent memory — Microsoft open-sources STATE-Bench, a memory-agnostic benchmark for evaluating AI agent memory across platforms. Outcome engineers should treat this as a standard test-suite for persistent agents: use it to validate long-term state, regression across context stores, and to drive auditability and recall guarantees (Principles 16, 14).

Constraint Decay: The Fragility of LLM Agents in Backend Code Generation — the paper demonstrates that LLM agents increasingly fail to enforce multi-file architectural constraints, with correctness collapsing as requirements accumulate. Build teams need to bake structural verification, modular decomposition, and multi-file unit tests into agent workflows rather than assuming single-pass LLM correctness (Principles 14, 16).

Pi Demonstrates Self-Modifying AI Coding Agent — Pi shows minimalist self-modifying coding agents that change their own code and workflows while emphasizing verification and human oversight. If you consider self-modifying agents, invest in immutable audit trails, human-in-the-loop gates, and executable proofs of change before deployment (Principles 14, 15, 16).

Hadrian releases OpenHack for AI vulnerability research — Hadrian open-sources OpenHack, a file-backed AI code-review workflow that scopes scenarios, separates triage, and preserves artifacts to reduce hallucinations. Adopt its file-backed, artifact-first pattern to maintain reproducible reviews, preserve evidence for audits, and limit attack surface during automated reviews (Principles 02, 14).

The role of MCP in context engineering — the piece describes the Model Context Protocol (MCP) for standardizing real-time connections between agents and data sources to unlock scalable context engineering. Make MCP-style interfaces the center of your context layer so agents get reliable, versioned context feeds that support traceability and safe updates (Principles 06, 11).