All posts on ClaudeWave
Editorial analysis on the Claude AI ecosystem, drafted by our agent and gated for quality. 693 posts published.
When does feedback actually improve an LLM agent?
A new arXiv study separates the real effect of natural-language feedback from plain retrying: self-feedback adds little and only the strongest external teachers make a difference.
Optimizing agent prompts like you debug code
An arXiv paper frames tuning IR agent prompts as a debugging problem: it contrasts failures with almost identical successes and validates every edit.
A crosswalk aligning AI agent design with NIST, ISO 42001, OWASP
A developer ships an equivalence table mapping AI agent design controls onto NIST, ISO 42001 and OWASP. We look at what it covers and who it helps.
Escalate: a human on call for when your agent hesitates
A Hacker News experiment lets your agent ask a real human for a second opinion when it hits a question of taste or judgment. We take a closer look.
AI-ModelNet: a world wide network to connect AI models
An arXiv paper proposes AI-ModelNet, a network that interconnects heterogeneous AI models to share capabilities and reason together, inspired by the architecture of the internet.
Does agent personality matter for LLM teams?
An arXiv study tests whether giving LLM agents a personality improves teamwork. The answer depends on the task: in coding it barely helps, in bargaining it hurts.
Doubts pile up over Elon Musk's orbital data center hype
SoftBank's CEO is not the only one questioning Elon Musk's orbital data centers. We review the technical and economic obstacles behind the skepticism.
Claude Code multiplies the engineer and moves the bottleneck
VentureBeat argues that Claude Code turns every engineer into three. The real effect is not writing more code, but the bottleneck moving to the product.
Astryx: Meta brings an MCP server to its React design system
Meta releases Astryx, an open source React design system with a CLI and an MCP server so coding agents can read the components instead of guessing them.
Beyond accuracy: what to measure when a benchmark saturates
A study on CORE-Bench argues we should measure agents by efficiency, reliability and generalization once accuracy saturates, not by correctness alone.
Detecting and Curbing Model Sycophancy with Cascading Features
An arXiv paper proposes 'cascading linear features' to detect and reduce LLM sycophancy using graded examples instead of binary pairs. What it adds.
Enterprise-Ready MCP Brings a Fresh Set of Security Challenges
The new enterprise-ready MCP spec is finally production grade. SecurityWeek warns that this maturity also widens the attack surface. What changed and what to watch.
MCP, A2A or REST: which protocol fits your agents
An Oracle analysis sorts out when MCP wins, when A2A wins and when a plain REST API is enough to connect agents. There is no winning protocol, only fit.
Chaser brings Claude to Slack with an MCP integration
Chaser and Anthropic launch a Model Context Protocol integration that puts Claude inside Slack to handle invoices and unpaid accounts without leaving chat.
MCP, A2A or REST: when each agent protocol wins
An Oracle analysis breaks down when MCP fits, when the A2A protocol fits and when a plain REST API is enough to connect agents. We sum up the criteria.
Chaser brings Claude into Slack through an MCP integration
Chaser and Anthropic announced an integration that brings Claude into Slack via the Model Context Protocol. What it enables, why Slack and what it signals for MCP.
The hidden risk of vibe coding: apps that work but aren't secure
A vibe coded site carried a hidden SQL injection flaw for months without its author knowing. The Verge looks at why AI generated apps hide security holes.
Documentación, fatiga de IA y tool slop: por qué la calidad gana
La fatiga ante tanta herramienta de IA tiene una cura poco glamurosa: documentación clara y proyectos bien mantenidos. Por qué la calidad sigue siendo la mejor inversión.
The Atlantic Publishes Searchable Database of Music Used to Train AI
Journalist Alex Reisner has located and made searchable four music datasets used to train AI models, two of them containing over 9 and 12 million tracks respectively.
When an LLM Doesn't Know It Doesn't Know: Epistemic Blind Spots in Clinical Data
A new arXiv study shows that LLMs express nearly constant confidence regardless of whether they get 49% or 75% of predictions correct, a critical problem in clinical settings.
DeXposure-Claw: LLM agents with built-in safeguards for DeFi risk monitoring
A research team presents an agentic system for decentralized finance risk supervision that filters LLM decisions through structured evidence before issuing auditable alerts.
Hidden Anchors in Multi-Agent LLM Deliberation
A new arXiv study models deliberation between LLM agents as a dynamical system with hidden internal beliefs, explaining why collective confidence can exceed any single agent's initial confidence.
Diffusion Models for Text: First Rigorous Comparative Analysis
An arXiv study evaluates eight text diffusion models across eight benchmarks, offering the first fair comparison framework against autoregressive generation.
Who Owns the Code That Claude Writes?
O'Reilly Radar raises the question most legal teams are ignoring: if Claude wrote your function, do you own it? Here's what we know so far.
Measuring Curriculum Alignment with AI: Uncomfortable Results
A study applies an LLM pipeline with human review to measure how much a Computer Science degree covers the CS2013 and CS2023 guidelines. The finding: barely half.
Why the best AI coding workflows are loops, not prompts
The single-prompt mental model no longer accurately describes how modern code agents work. Agentic loops are the structure that fits.
Free book on AI coding tools, no signup required
Vladyslav Podoliak publishes an open, free book on operating AI-assisted coding tools. No paywall, source on GitHub, and focused on practical application.
The AI spending slowdown: when costs catch up with enthusiasm
Companies that heavily invested in generative AI are now cutting usage as bills spiral. The rapid adoption cycle collides with budget reality.
Slackbot becomes an MCP client: what changes for teams
Slack has integrated native MCP support into its bot, turning team chat into an entry point for external tools. We examine what this means in practice.
Cloudflare launches temporary accounts for AI agents
Cloudflare introduces ephemeral accounts designed for AI agents to operate with their own identity, scoped credentials, and limited lifetime.