Tag
AWS launches Amazon CloudWatch Omni to provide comprehensive observability for AI agents and generative AI workloads, emphasizing the need to measure correctness alongside traditional metrics.
The article explores the challenges of delegation and permission inheritance in agentic systems, raising questions about autonomy and authority management when agents spawn sub-agents.
The author advocates for designing AI agents that enhance human thinking and creativity without replacement, emphasizing the need to preserve problem-solving and learning in an agentic environment.
This article presents a simple test to distinguish between AI agents and traditional workflows based on runtime decision dependency, discusses architectural trade-offs, and highlights common production failures in agentic systems.
This article summarizes Ion Stoica's talk at Ray Summit, exploring the three key gaps in requirements, environment, and evaluation faced by AI programming agents in software engineering, and how these issues lead to reward hacking and hallucinations.
OpenAI's report on six new misalignment cases highlights the critical need for local sandboxes in AI safety, arguing that central model guardrails are insufficient and infrastructure-level safety is essential.
This post provides a guide for setting up agentic systems, emphasizing the importance of foundational tools like Linux, Tailscale, and specific engines before selecting an AI model. It outlines eight key components and habits for effective local and cloud deployment.
This paper reports on a production migration for a customer support conversational assistant, replacing a monolithic AI model with an agentic orchestration system to improve precision, reduce hallucinations, and lower serving costs.
Undergraduate research intern Matheus Kunzler Maldaner explores agentic systems and shares his experience in the first episode of Microsoft Research's series on research careers.
The author shares lessons learned from building agent graphs, highlighting the pitfalls of parallelism in cyclic workflows and advocating for sequential dependent checks and mechanisms to challenge feedback for better efficiency.
Uber details their 'Software Factory' vision, where AI agents manage over 70% of pull requests and achieve significant cost reductions through optimized AI usage across the software development lifecycle.
This survey defines agentic artifact creation as stateful, feedback-driven construction of deliverables by AI systems and analyzes 259 works to propose principles for accountable control.
LLM4LLM introduces a deployment-aware closed-loop optimization framework to bridge kernel benchmarks and real LLM inference, achieving up to 6.98x speedups on H100 GPUs.
A tweet from @cyrilXBT praises a paper on graph engineering for agentic systems, claiming it significantly improves AI responses, and promotes a spammy bidding website called RankNow.
Andrew Ng has released the latest AI Engineer skill map, detailing six core competencies for building and deploying AI applications, and emphasizing the importance of the Eval loop in AI system engineering.
This article outlines the key skills required for building and deploying AI applications, emphasizing LLM foundations, data grounding, and agentic systems, based on analysis of job postings and expert interviews.
MEGA is a four-week intensive course starting October 26, designed to help developers master AI and build agentic systems through live workshops with industry experts.
The article discusses the importance of system design for AI agents, covering concepts like Agent Harness, LLMOps, and Evals, and provides a proof-of-concept implementation with plans for future parts.
The paper presents AgentChaosBench, a benchmark for detecting and localizing runtime faults in LLM-based agentic systems, and evaluates it using zero-shot LLM baselines, revealing significant challenges in fault diagnosis.
This paper evaluates four open-source LLM routers using a common protocol across four benchmarks, finding that performance gains are more closely tied to model tier composition than task-specific targeting.