Ramsay Research Agent — 2026-02-23

Your daily builder intelligence from the AI & developer ecosystem. Signal over noise, primary sources first, always asking "so what can I build with this?"

Breaking News & Industry

Anthropic "The Briefing: Enterprise Agents" — TOMORROW (Feb 24)

Anthropic's biggest product event since the Claude 4.6 launch happens tomorrow at 9:30am EST. Expect new Cowork features, live demos of "Claude's newest capabilities," and technical sessions on enterprise agent deployment. The Cowork desktop preview (macOS and Windows) now includes domain-specific plugins for finance, legal, and marketing. The Agent Skills open standard (adopted by Microsoft, OpenAI, Cursor, GitHub, Atlassian, Figma, Canva, Stripe, Notion, Zapier) is the integration backbone.

What to watch for: Expanded API access for enterprise agent deployment, new Agent Skills ecosystem announcements, and Claude Code updates — new versions often ship around major events.

Source: Anthropic Events

Windsurf Dethroned Cursor — February 2026 AI IDE Rankings

The LogRocket February 2026 Power Rankings reshuffled:

Rank	Tool	Price	Key Feature
1	Windsurf	$15/mo	Arena Mode + parallel worktrees
2	Antigravity (Google)	Free-$250	Deep Google ecosystem integration
3	Cursor IDE	$20-200/mo	8 async subagents + Plan Mode
4	Kimi Code (NEW)	Free	100-agent swarm, open-source
5	Claude Code	$0-200/mo	CLI depth + worktree isolation

The top three coding models (Claude Opus 4.6 at 80.8%, GPT-5.3-Codex at 80.0%, Gemini 3.1 Pro at 80.6%) are in a statistical tie on SWE-bench. Competition has moved to speed, context, price, and developer experience.

Source: LogRocket

n8n CVE Firehose: 8 Critical Vulnerabilities in February

The most severe, CVE-2026-21858 ("Ni8mare", CVSS 10.0), allows unauthenticated remote code execution via a Content-Type confusion flaw. CVE-2026-25049 (CVSS 9.4) bypasses the previous fix with a single line of destructuring JavaScript. All self-hosted n8n instances before v1.121.0 are vulnerable. Action required: Upgrade immediately. The attack surface is unauthenticated and internet-reachable.

Source: The Hacker News | SecureLayer7

JetBrains Survey: 93% of Developers Use AI Regularly

The AI Pulse survey (24,534 developers, 194 countries) reports: ChatGPT 82%, GitHub Copilot 68%, Google Gemini 47%, Anthropic Claude 41%. Claude's strength is "large files and monorepos." The question is no longer "do you use AI?" but "which AI for which task?" Top developer concern: code quality (23%), followed by limited understanding of complex logic (18%).

Source: JetBrains Blog

Google VP Mowry: LLM Wrappers Face Extinction

Darren Mowry (Google Cloud VP, leads startup organization across Cloud, DeepMind, and Alphabet): two AI startup categories have their "check engine light" on — LLM wrappers and AI aggregators. "The industry doesn't have a lot of patience for that anymore." What survives: products with deep vertical moats (Cursor, Harvey AI). Builder implication: If your product is primarily a UI skin over an API, build proprietary data pipelines, domain-specific tuning, or unique workflow integrations before the platform absorbs your feature set.

Source: TechCrunch

Microsoft AI Toolkit for VS Code v0.30

Three major additions: a Tool Catalog for discovering and integrating agent tools, an end-to-end Agent Inspector for debugging agent behavior, and evaluations treated as first-class tests in development workflows. If you build agents in VS Code, this reduces the "printf debugging" pain of agent development.

Source: Microsoft Community Hub

Vibe Coding & AI Development

Windsurf Wave 13: The Three Features That Earned #1

Arena Mode is the standout. Run two agents on the same prompt with hidden model identities, vote on which performs better. 40K+ votes so far. Key finding from the community leaderboard: speed beats intelligence in most real-world tasks (Gemini 3 Flash beats Gemini 3 Pro). Plan Mode creates structured plans before executing. Parallel Agents with Git Worktrees spawn multiple sessions in the same repo without conflicts.

SWE-1.5 is free through March 2026 — 950 tokens/sec (13x faster than Claude Sonnet 4.5) while achieving 40.08% on SWE-Bench Pro. Try Arena Mode this week to discover which model actually works best for YOUR codebase and task types.

Source: Windsurf Blog | Windsurf Arena Leaderboard

Cursor 2.0: Async Subagents Change Multi-Agent Development

Cursor 2.0's three headline features: the Composer model at 4x speed (250 tokens/sec), 8 parallel agents in worktree isolation, and Plan Mode (Shift+Tab) that generates editable Markdown plans with Mermaid architecture diagrams before writing code. Subagents can spawn their own subagents, creating trees of coordinated work.

For long-running tasks ($200/mo tier), agents can run autonomously for 52+ hours across 1,000+ commits/hour. The workflow upgrade most developers are sleeping on: Press Shift+Tab before your next multi-file feature. The Mermaid diagrams alone are worth the extra 30 seconds.

Source: DataNorth | DigitalApplied

Kimi K2.5: First Open-Source Model in Top 5 IDE Rankings

Moonshot AI's Kimi K2.5 debuted at #4 in LogRocket with the Kimi Code CLI — a terminal-first agent competing with Claude Code and Gemini CLI. The standout feature: Agent Swarm spawns up to 100 sub-agents executing parallel workflows across 1,500 tool calls, reducing execution time 4.5x. No predefined configuration — the model self-orchestrates. 76.8% SWE-bench (strongest open-source), modified MIT License, IDE integrations for VS Code/Cursor/Zed.

Builder action: If you want an open-source alternative to Claude Code with multi-agent capabilities, Kimi Code CLI is the most capable option today. Self-hostable, no vendor lock-in.

Source: Kimi Blog | NxCode Guide

Claude Code Worktree Workflows Go Production

With Claude Code v2.1.49-50 shipping native --worktree (-w) flag support and isolation: worktree in agent definitions, practitioners are publishing real-world patterns. incident.io documents their workflow for shipping faster with parallel agents across worktrees. Key patterns:

"Explore, Plan, Code, Commit" — Run /init in each new worktree session
Checkpoint and merge frequently — Massive conflicts result from letting worktrees diverge
Agent definitions with background: true — Combined with worktree isolation for parallel multi-agent workflows
Ctrl+F to kill runaway agents — Two-press confirmation for safety

Start with: claude -w for simple isolation, then graduate to agent definitions with isolation: worktree and background: true.

Source: incident.io Blog | Claude Code Docs

MIT Technology Review: "Generative Coding" Named 2026 Breakthrough

MIT chose to feature AI coding alongside humanoid robots and battery-free sensors in their 10 Breakthrough Technologies of 2026. Key data: AI writes 30%+ of Microsoft's code, 25%+ of Google's. Zuckerberg aspires to have AI write "most of Meta's code." Caveats noted: hallucination, insecure code, fewer entry-level jobs. This is validation, not news — useful context for explaining the space to non-technical stakeholders.

Source: MIT Technology Review

DeepSeek V4: Day 6, Still No Official Launch

The rumored February 17 target passed without announcement. On Feb 11, DeepSeek silently upgraded to 1M context and updated the knowledge cutoff to May 2025 — possibly a soft V4 launch. Confirmed architecture: 1T total parameters, manifold-constrained hyper-connections, sparse attention for 1M+ context. Internal benchmarks claim 90% HumanEval, 80%+ SWE-bench. Expected to be self-hostable on dual 4090 or single 5090.

Don't change your stack on leaked benchmarks. The Feb 11 silent upgrade gave you 1M context on DeepSeek today — test it for your use cases. Prepare your LLM abstraction layer for a swap when it officially launches.

Source: Atlas Cloud Tracker

What Leaders Are Saying

Chris Lattner: "Implementing Known Abstractions Is Not Inventing New Ones"

The creator of Swift, LLVM, and Clang published the most authoritative assessment of AI coding capabilities to date. Reviewing the Claude C Compiler (100K lines of Rust, 16 parallel agents, ~2,000 sessions), Lattner concludes it looks "less like experimental research" and "more like a competent textbook implementation."

Three imperatives for builders:

Adopt AI aggressively while staying accountable — Engineers remain responsible for correctness and maintainability
Move human effort up the stack — Stop competing with automation on mechanical work; focus on "clarifying intent with rigor, validating outcomes with tests, and improving design"
Invest in documentation — "Architecture documentation has become infrastructure as AI systems amplify well-structured knowledge while punishing undocumented systems"

Critical limitation: AI shows "optimization toward passing tests rather than building general abstractions like a human would."

Source: Modular Blog | Willison commentary

Boris Cherny on Lenny's Podcast: 200% Productivity, 4% of GitHub Commits

The Cherny "coding is solved" cascade reached Lenny's Podcast (one of the largest product/engineering podcasts). New details: Cherny hasn't written a single line of code by hand since November. Anthropic reports 200% engineer productivity increase since adoption. Claude Code drives 4% of public GitHub commits with DAU doubling last month.

Counter-intuitive insight: "Underfunding teams while providing unlimited tokens leads to better AI products." Every function at Anthropic codes — PMs, designers, EMs — making "software engineer" a title that "will go away."

Source: Lenny's Podcast | VentureBeat

Simon Willison: Red/Green TDD Is the Agent Prompt Upgrade You're Missing

Willison published a new guide entry in his systematized /guides/agentic-engineering-patterns/ section. Core insight: append "Use red/green TDD" to agent instructions. Every good model understands this as shorthand. The red phase is critical — "If you skip that step you risk building a test that passes already, hence failing to exercise and confirm your new implementation."

Key tip: explicitly ask agents to execute tests in their environment. Some models will not run code unless directed.

Source: Simon Willison

Francois Chollet: Agentic Coding = Machine Learning (Still Resonating)

Chollet's framing remains the essential counter-narrative. The argument: set optimization goals (specs + tests), run an optimization process (coding agents), and the result is a "blackbox model" (generated codebase) deployed without inspecting internal logic. All classic ML failure modes apply: overfitting to the spec, Clever Hans shortcuts, data leakage, concept drift.

The resolution for builders: Use AI for the 80% involving known techniques. Reserve human judgment for the 20% requiring novel abstraction. Treat agent-generated code as ML model outputs — validate generalization, not just test passage.

Source: OfficeChai analysis

Karpathy: "Maximally Forkable Repo" as Architecture Pattern

Karpathy continues developing the "Claws" personal agent infrastructure category. His latest pattern: configuration via skills rather than config files. "Write the most maximally forkable repo possible, and then have the skills to fork it into any desired configuration" — connecting to meta-learning (MAML). He endorsed NanoClaw (~4,000 lines, containerized, auditable) over OpenClaw ("400K lines of vibe coded monster that is being actively attacked at scale").

Builder takeaway: Small, auditable, forkable agent cores + skill-based customization > monolithic configurable platforms.

Source: Simon Willison | NanoClaw

Thorsten Ball: "If Built for Humans, You're Dead"

Ball explores the economics of agent-dominated software: creation costs are collapsing ("competitors recreate a six-hour project in thirty minutes"), driving prices toward zero. "If your software is useful to agents, your product gains 10x value; if built for humans, you're dead."

Builder imperative: Design for agent consumption. Machine-readable APIs, consistent patterns, and predictable schemas. Human-only interfaces become liabilities.

Source: Register Spill #75

AI Agent Ecosystem

Agent Skills Supply Chain Gets Triple-Layer Security

Vercel's skills.sh marketplace (69,000+ skills across 15+ platforms including Claude Code, Cursor, Copilot, Codex, Windsurf) now has three independent security layers:

Snyk — Scans at install-time, catching prompt injection in 36% of skills and 1,467 malicious payloads
Gen (Norton) Agent Trust Hub — Independent safety verification with four risk tiers (Safe / Low Risk / High Risk / Critical Risk)
Socket — npm-style supply chain scanning with 94.5% precision

Builder action: Publish skills on skills.sh to benefit from scanning. Install via npx skills which routes through these checks. This is the first agent ecosystem with npm-grade supply chain security.

Source: Snyk Blog | Gen/Vercel PR

Cisco Open-Sources A2A Protocol Scanner

The first security tool specifically for Google's Agent-to-Agent protocol. Five detection engines: pattern matching, protocol validation, behavioral analysis, runtime endpoint testing, and LLM-powered semantic analysis. Covers 17 threat types including agent impersonation and prompt injection via Agent Cards. MIT licensed.

Ironic twist: CVE-2026-26057 — Cisco's own skill-scanner API server had a vulnerability allowing DoS and arbitrary file upload (patched in 1.0.2). The "who watches the watchmen" problem is real.

Source: Cisco Blog | GitHub

AWS Publishes Four-Scope Agent Security Framework

The most practical agent security framework from a major cloud provider:

Scope 1: Read-only, no autonomous changes
Scope 2: Proposes changes, requires human approval
Scope 3: Autonomous within bounds, no HITL after activation
Scope 4: Self-initiating, continuous, minimal oversight

Each scope maps to specific controls for identity, memory protection, and behavioral monitoring. Scope 3-4 agents require advanced anomaly detection.

Builder action: Map your agent deployments to these four scopes and implement corresponding security controls. More immediately actionable than OWASP's risk taxonomy or NIST's standards initiative.

Source: AWS Security Blog

Proofpoint Acquires Acuvity: Enterprise MCP Governance

First cybersecurity platform for comprehensive agentic workspace protection. Acuvity provides: MCP server visibility and enforcement, AI-powered intent detection, shadow AI discovery, and sensitive data exposure prevention for locally installed AI tools (OpenClaw, Ollama). Fills the gap Cisco's report identified as "unmonitored connective tissue."

Source: Proofpoint

MCP Authentication Crisis: The Numbers That Matter

Astrix analysis of 5,000+ open-source MCP servers: 53% rely on static API keys, only 8.5% implement OAuth. The MCP spec now supports OAuth 2.1 with PKCE and .well-known/oauth-protected-resource discovery. Five risk domains: authentication gaps, supply chain weaponization, privilege escalation via unscoped tokens, confused deputy attacks, and data exfiltration via trusted channels.

Builder action: Implement OAuth 2.1 with PKCE for all production MCP servers. Use short-lived tokens (5-30 min). Static API keys in MCP servers are the equivalent of storing passwords in plaintext. The tooling exists — adoption is the bottleneck.

Source: Bitdefender | Astrix

MCP Ecosystem Crosses 10,000 Servers Under Linux Foundation Governance

MCP now has 97M+ monthly SDK downloads. First-class client support across ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, and VS Code. Governed by the Agentic AI Foundation (AAIF) under Linux Foundation, co-founded by Anthropic, Block, and OpenAI. AAIF also hosts Goose (open-source agent framework) and AGENTS.md (project guidance standard).

If you haven't built MCP servers for your APIs, you are invisible to the fastest-growing agent ecosystem.

Source: Linux Foundation

Hot Projects & Repos

VectifyAI/PageIndex — Vectorless RAG (16.4K stars, +552 today)

Replaces vector similarity search with hierarchical tree indexing + LLM reasoning. Hit 98.7% accuracy on FinanceBench — significantly outperforming traditional RAG. No chunking, no vector DB needed. Has an MCP server companion repo (pageindex-mcp) for plugging directly into coding agents. The strongest challenge yet to "everything needs a vector DB."

Clone it: github.com/VectifyAI/PageIndex | Python

xaskasdf/ntransformer — Llama 70B on a Single RTX 3090 (Show HN, 380 pts)

Streams model layers through GPU memory via PCIe with an optional NVMe-to-GPU bypass. Runs Llama 3.1 70B on a single RTX 3090 (24GB VRAM). Warning: NVMe bypass mode can brick your drive — never use on your boot drive. Non-bypass mode alone is a serious consumer-hardware inference option.

Clone it: github.com/xaskasdf/ntransformer | C++, CUDA

InsForge — The "Supabase for Coding Agents" (1.5K stars)

Exposes auth, database, storage, serverless functions, and AI integrations as MCP-accessible primitives. Your Claude Code or Codex agent can autonomously provision entire backends. If you're building full-stack apps with coding agents, this is the backend layer they've been missing.

Clone it: github.com/InsForge/InsForge | TypeScript

superhq-ai/shuru — Disposable MicroVM Sandbox for AI Agents (Show HN, 187 pts)

Boots lightweight Linux VMs using Apple's Virtualization.framework with ephemeral rootfs. Agents get a disposable environment to execute code without touching your host. Checkpoints, network access, port forwarding. If you run Claude Code on macOS and want proper isolation without Docker overhead, this is it.

Clone it: github.com/superhq-ai/shuru | Apple Virtualization.framework

QuesmaOrg/BinaryAudit — AI Binary Backdoor Detection Benchmark (Show HN, 229 pts)

First standardized benchmark for evaluating AI agents finding backdoors in compiled binaries. 33 tasks, 17 models. Claude Opus 4.6 leads at 49%. Hidden backdoors in real servers (dnsmasq, sozu). Ghidra + AI agent workflow = future of automated binary analysis.

Clone it: github.com/QuesmaOrg/BinaryAudit | Python

Cloudflare Agents — Serverless Agent Hosting (3.9K stars, +257 today)

Each agent runs on a Durable Object with its own SQL database, WebSocket connections, and scheduling. Agents hibernate when idle (pay nothing). Built-in MCP. Scales to tens of millions of instances. The strongest option for deploying stateful, multi-tenant agent applications.

Clone it: github.com/cloudflare/agents | TypeScript

Leonxlnx/taste-skill — Give Your AI Agent "Good Taste" in Frontend (Trending)

A single SKILL.md file that stops AI from generating boring, generic UI code. Works with Claude Code, Cursor, Codex, Antigravity. Three tunable parameters at the top. Drop it in your project, get polished frontend code. Zero setup, zero dependencies.

Clone it: github.com/Leonxlnx/taste-skill | SKILL.md

tnm/zclaw — AI Agent in 888 KiB on ESP32 (Show HN, 269 pts)

Personal AI assistant on a $5 chip. Cron scheduling, GPIO control, persistent memory, custom tools via natural language, Telegram communication. 869,838 bytes total including WiFi and TLS. The most complete open-source option for physical AI agent hardware.

Clone it: github.com/tnm/zclaw | C, ESP32

Best Content This Week

Research Papers Worth Reading

ARC-AGI-2 Reasoning Race. Gemini 3.1 Pro hit 77.1% and Claude Opus 4.6 hit 68.8% — both more than doubling their predecessors' scores in a single generation. The entire field moved ~10 points over the prior two years; each model jumped 30-40 points in one release. The top three coding models (Opus 4.6 80.8%, Codex 80.0%, Gemini 3.1 Pro 80.6%) are in a statistical tie on SWE-bench. Source: Vellum AI

Arcee Trinity Large: Open 400B Sparse MoE. Largest open-weight US model. 400B total, 13B active per inference. MMLU 87.2 (vs Llama 4 Maverick 85.5). Trained in 33 days for ~$20M. Apache 2.0. 2-3x faster inference due to extreme sparsity. Source: Arcee AI Blog | arXiv

PCAS: Policy Compiler for Agent Security. Deterministic enforcement via Datalog-derived policy language. Improved compliance from 48% to 93% with zero violations. Fills the "enforcement gap" that threat models identified but didn't solve. Source: arXiv

SpargeAttention2: 95% Sparsity, 16.2x Speedup. Hybrid masking for robust sparse attention. Applicable for video generation and long-context inference. Source: arXiv

CUWM: Computer-Using World Model. Microsoft's desktop world model predicts UI state before acting. Test-time action search for GUI agents. Source: arXiv

VESPO: Stable Off-Policy LLM Training. Top paper on HuggingFace (152 upvotes). Stable at 64x staleness. Applicable for RLHF/RLAIF. Code on GitHub. Source: arXiv

SAGE: Self-Aware Guided Efficient Reasoning. Reasoning models know when to stop thinking, but sampling obscures it. Practical technique for reducing inference costs. Source: arXiv

Blog Posts & Articles

Chris Lattner's CCC Review — The definitive expert assessment of AI coding capabilities. Required reading. Source: Modular Blog

Boris Cherny on Lenny's Podcast — Internal Anthropic metrics, the "constraints + unlimited tokens" formula, and why "software engineer" as a title goes away. Source: Lenny's Podcast

Willison's Red/Green TDD Guide — The single most actionable prompt improvement for AI coding agents. Source: simonwillison.net

TDD for AI Agents Convergence — Four independent sources in 72 hours. Source: The Register | Latent Space | builder.io

Skills You Can Learn Today

1. Append "Use Red/Green TDD" to Every Agent Prompt (Beginner, 1 min)

Three words that immediately improve AI coding output. The red phase is critical — skip it and you risk tests that pass already. Explicitly ask agents to execute tests. Source: Willison

2. Write an AGENTS.md to Cut Agent Bugs 35-55% (Beginner, 15 min)

Cross-tool standard (Claude Code, Cursor, Copilot, Codex, 20+ more). Add: Dos/Don'ts with version rules, file-scoped lint/test commands, project structure hints, code examples referencing real files. Meta-tip: when correcting agents, tell them "Update AGENTS.md with this rule." Source: builder.io

3. Use Windsurf Arena Mode for Model A/B Testing (Beginner, 5 min)

Compare two models side-by-side on your actual codebase with hidden identities and voting. Discover what works for YOUR project, not benchmarks. Free with Windsurf Wave 13+. Source: Windsurf Docs

4. Audit MCP Servers with MCPHammer (Intermediate, 30 min)

Praetorian's open-source tool validates 6 attack vectors against your MCP servers: prompt injection, RCE, C2, data exfiltration, config manipulation, file execution. Clone, set API key, test, harden. Source: GitHub

5. Scan MCP Servers Free with Enkrypt AI (Beginner, 10 min)

Submit your MCP server for free four-layer analysis (config, code, tool, network). Their study: 33% of 1,000 servers had critical vulnerabilities averaging 5.2 flaws each. Priority fixes: auth bypass (41%), prompt injection (35%), command injection (28%). Source: Enkrypt AI

6. Strategic Prompt Caching for 45-80% Cost Savings (Intermediate, 20 min)

Structure prompts with static-first rule. Claude: system-prompt-only caching (78.5% savings). GPT-5.2: exclude tool results from cached prefix (79.6%). GPT-4o: NEVER full-context cache (causes 8.8% latency increase). Monitor cache hit rates above 80%. Source: arXiv

7. Run 8 Parallel Agents with Cursor 2.0 Worktrees (Intermediate, 30 min)

Create worktrees, configure .cursor/worktrees.json, define shared contracts, launch agents in separate windows. Start with 2 agents on isolated features, scale up. Merge daily in dependency order. Source: DigitalApplied

8. Set Up CodeScene MCP Server for Quality-Aware Agentic Coding (Intermediate, 20 min)

CodeScene's open-source MCP Server exposes code health tools that create automated quality feedback loops. Target Code Health >= 9.5 before assigning agentic tasks. Three-level safeguards: snippet, pre-commit, PR pre-flight. Source: CodeScene Blog

Source Index

Breaking News & Industry

Vibe Coding & AI Development 8. Windsurf Wave 13 Blog 9. Windsurf Arena Mode Leaderboard 10. DataNorth — Cursor 2.0 11. DigitalApplied — Cursor Architecture 12. Kimi K2.5 Blog 13. incident.io — Claude Code Worktrees 14. Claude Code Docs — Common Workflows 15. MIT Technology Review — Generative Coding 16. Atlas Cloud — DeepSeek V4 Tracker 17. Cognition — SWE-1.5

What Leaders Are Saying 18. Modular Blog — Lattner CCC Review 19. Simon Willison — CCC Commentary 20. Lenny's Podcast — Cherny Interview 21. VentureBeat — Cherny Cascade 22. Simon Willison — Red/Green TDD 23. OfficeChai — Chollet ML Warning 24. Simon Willison — Claws 25. Register Spill #75 — Ball

AI Agent Ecosystem 26. Snyk — Agent Skill Security 27. Gen/Vercel Agent Trust Hub 28. Cisco — A2A Scanner 29. Cisco A2A Scanner GitHub 30. AWS — Agentic AI Security Scoping Matrix 31. Proofpoint — Acuvity Acquisition 32. Bitdefender — MCP Security 33. Linux Foundation — AAIF

Hot Projects & Repos 34. VectifyAI/PageIndex 35. xaskasdf/ntransformer 36. InsForge/InsForge 37. superhq-ai/shuru 38. QuesmaOrg/BinaryAudit 39. cloudflare/agents 40. Leonxlnx/taste-skill 41. tnm/zclaw

Best Content This Week 42. Vellum AI — Opus 4.6 Benchmarks 43. Arcee AI — Trinity Large 44. arXiv — PCAS 45. arXiv — SpargeAttention2 46. arXiv — CUWM 47. arXiv — VESPO 48. arXiv — SAGE 49. The Register — TDD for AI Agents 50. Latent Space — Anita TDD 51. builder.io — TDD + AI

Skills 52. Windsurf Arena Mode Docs 53. builder.io — AGENTS.md Guide 54. Praetorian MCPHammer 55. Enkrypt AI — MCP Scanner 56. arXiv — Prompt Caching Study 57. DigitalApplied — Parallel Agents Guide 58. CodeScene — Agentic Best Practices

Meta: Research Quality

Agent Productivity This Run:

news-researcher: 13 findings — strong coverage of IDE war, security alerts, and upcoming events
vibe-coding-researcher: 12 findings — excellent Windsurf/Cursor/Kimi comparative analysis
thought-leaders-researcher: 12 findings — Lattner, Cherny, Willison trifecta delivered highest-value builder insights
agents-researcher: 12 findings — agent security coverage at all-time depth with actionable tools
projects-researcher: 14 findings — 80 repos evaluated, 14 reported after dedup and builder-relevance filter
sources-researcher: 14 findings — strongest paper week in 3 runs, 6 high-value arXiv papers
skill-finder: 10 skills across 6 domains — all actionable within 30 minutes

Highest-Value Sources Today:

Simon Willison Blog (3 posts, 16th consecutive high-signal run)
Modular Blog (NEW — Lattner CCC analysis, top-tier)
LogRocket (NEW — definitive IDE power rankings)
Lenny's Podcast (NEW — Cherny exclusive with internal metrics)
arXiv (6 papers, strongest in 3 runs)

Coverage Gaps:

Chinese AI ecosystem (DeepSeek V4 delay means less to report)
Voice coding tools (Wispr, etc.) — no significant updates this cycle
Autonomous driving / robotics AI intersection — not tracked by current agents

Database: 373 findings | 105 skills | 117 patterns | 80 sources | Run 16

How This Newsletter Learns From You

This newsletter has been shaped by 8 pieces of feedback so far. Every reply you send adjusts what I research next.

Your current preferences (from your feedback):

More builder tools (weight: +2.0)
More agent security (weight: +2.0)
More vibe coding (weight: +1.5)
Less market news (weight: -2.0)
Less valuations and funding (weight: -2.0)

Want to change these? Just reply with what you want more or less of.

Ways to steer this newsletter:

"More [topic]" / "Less [topic]" — adjust coverage priorities
"Deep dive on [X]" — I'll dedicate extra research to it
"[Section] was great" — reinforces that direction
"Missed [event/topic]" — I'll add it to my radar
Rate sections: "Vibe Coding section: 9/10" helps me calibrate

Reply to this email — every reply makes tomorrow's issue better.