The Infrastructure Layer Hardens: Agent Runtimes, Security Models, and the Case Against Premature Abstraction

Tools

LangGraph Ships Production Session Management

This solves one of the most annoying production problems: keeping agent state alive across user sessions without rebuilding context from scratch. If you're deploying agents that need to remember conversations, this is the pattern to study.

MCP Specification Updates and Integration Examples

The MCP docs continue to expand with real integration patterns, not just theoretical protocols. If you're building tool-using agents, this is becoming the reference implementation to benchmark against.

Research

Formal Verification Methods Applied to Agent Tool Invocation

Researchers are applying formal methods to verify that agents call tools correctly before execution. This is the kind of boring-but-critical work that separates toy demos from systems you'd trust with your API keys.

Sandbox Architectures for Untrusted Agent Code Execution

New paper on lightweight sandboxing for agents that write and run code. The key insight: you don't need full VMs if you get the syscall filtering right. Worth reading if you're building anything where agents touch real systems.

Cost Prediction Models for Multi-Step Agent Workflows

New research on predicting total agent cost before execution, not just tracking it after. This matters because most teams are flying blind on spend until the invoice lands. The models aren't perfect but they're directionally useful.

Analysis

Failure Mode Taxonomy for Autonomous Agent Systems

Comprehensive breakdown of how agents actually fail in production, not in benchmarks. The categories are practical: silent tool failures, context window exhaustion, and the classic 'agent confidently does the wrong thing' loop. Read this before your first real deployment.