AutoAgent Beats Human-Tuned Benchmarks and the Governance Stack Scrambles to Keep Up

Tools

AutoAgent Hits #1 on Two Benchmarks After Optimizing Itself for 24 Hours

This is the first public evidence of an AI agent autonomously beating hand-tuned harnesses on production benchmarks, hitting 96.5% on SpreadsheetBench and 55.1% on TerminalBench. The architecture is dead simple: a meta-agent edits the task agent's code, runs evals in parallel Docker sandboxes, and keeps only the changes that improve the score. If you're still hand-tuning agent harnesses, this is the writing on the wall.

Microsoft Open-Sources Agent Governance Toolkit with Sub-Millisecond Policy Enforcement

Microsoft's Agent Governance Toolkit is a seven-package, multi-language runtime that intercepts every agent action before it executes, covering all 10 OWASP agentic AI risks. The kernel-for-agents framing is apt and the sub-0.1ms p99 latency makes it viable for production use without adding meaningful overhead.

Analysis

AI Agent Governance is a Runtime Problem, Not a Dashboard Problem

This piece draws a sharp distinction that most teams miss: observability tells you what happened, governance decides what should happen. The stat that only 46% of organizations have an AI governance framework in place, while 64% of large companies have already lost over $1M to AI failures, makes the cost of waiting concrete.

Research

SMAS: A Formal Architecture for Agents That Rewrite Their Own Tools and Workflows

MARIA OS introduces a Self-Modifying Agent System with Lyapunov stability constraints and a defined modification frontier, meaning agents can update their own tooling without diverging into instability. The immutable audit trail and rollback capability are the practical bits that make this deployable rather than just theoretical.

Tutorials

Tutorial: Build a Meta-Agent That Designs, Instantiates, and Refines Task-Specific Agents

A hands-on Colab-ready tutorial covering dynamic agent design automation, tool selection, memory strategy, and iterative self-improvement in one unified framework. Worth bookmarking if you're building agent pipelines and want to move beyond static templates.