Research Explainers 9 min read

Context Engineering Is the New Systems Layer

A broad survey argues that the hard part of useful AI systems is no longer prompt wording but deciding what information reaches the model, when, and under what constraints. That makes retrieval, memory, permissions, timing, and context assembly a systems discipline.
Research Explainers 10 min read

Agents Are Reasoning Loops, Not Smart Prompts

A broad survey frames agentic AI as reasoning through action: planning, tool use, search, memory, feedback, and coordination across time. The framework clarifies why runtime design matters as much as model intelligence.
Research Explainers 10 min read

Agent Skills Need a Trust Layer

A survey argues that agent skills are becoming the packaging layer for procedural AI work, making governance, permissions, and verification unavoidable. It maps the controls needed before reusable instructions can be trusted across teams and organizations.
Research Explainers 9 min read

The Agent Is the Whole System

A survey argues that agent quality is a system property spanning models, memory, tools, planners, verifiers, permissions, traces, and evaluation. It offers a practical architecture for diagnosing failures without blaming the base model by default.
Research Explainers 8 min read

Multi-Agent Systems Fail Like Organizations

A NeurIPS dataset paper finds that multi-agent LLM systems fail through role confusion, broken handoffs, and weak verification, not only weak models. Its taxonomy turns coordination breakdowns into observable failure modes teams can test and repair.
Research Explainers 9 min read

Skill Bloat Is the New Context Tax

A paper argues that agent skills need a build-time optimization pass because many reusable instruction files waste context and make agents worse. Its proposed compiler trims redundancy while preserving the instructions that actually improve task performance.
Profile 7 min read

Lessons from Adrian Cockcroft

Adrian Cockcroft shaped cloud infrastructure at Sun Microsystems, Netflix, and AWS, demonstrating that microservices and chaos engineering could operate at massive scale. His observations connect speed and resilience with a crucial systems constraint: human behavior.
Profile 7 min read

Lessons from Christina Wodtke

Christina Wodtke, a Stanford professor and former tech executive, reframes Objectives and Key Results as tools for autonomy and continuous learning. Her work joins goal alignment, information architecture, and visual thinking in the practical design of self-managing teams.

Choose your reading rhythm.

Start with a weekly briefing, add daily notes, or hear only when a durable essay or research update is ready.
You've successfully subscribed to Antoine Buteau
You've successfully subscribed to Antoine Buteau
Welcome back! You've successfully signed in.
Press ESC to close.