In a study of 400,000 Claude Code sessions, Anthropic found that domain expertise, rather than programming ability, is the primary factor in effective autonomous AI work. Success rates converged across different professional fields, suggesting that agentic coding enhances existing expertise.

地味に効くんですよね、これ。同じプロンプトを1日に何十回も投げてると、トークン消費が2倍近くになってる気がする。 レスポンスをキーにしてローカルにキャッシュしておくだけで、API費用が実測40〜50%落ちます。実装は辞書1つ、5行のコードで完了。 Just cache identical prompts locally — costs drop 40-50% in practice. One dict, 5 lines of code, done. #AI #APIコスト削減 #ChatGPT #LLM #Prompting #キャッシュ #AI自動化 #TechT…

42% of companies scrapped AI projects last year (up from 17%). Not a tech problem. A measurement problem. Only 18% of orgs track ROI on AI spend. The other 82% fly blind — then act surprised when budgets vanish. 1,400+ sessions. Every metric published. Unmeasured AI fails.

headroom compresses tool outputs, logs, and RAG chunks before they hit the model, then serves the original back on demand. Its README claims 60 to 95 percent fewer tokens for the same answers. Drops in as a library, proxy, or MCP server. github.com/chopratejas/...

GitHub - chopratejas/headroom: Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 60-95% fewer tokens, same answers. Library, proxy, MCP server.

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 60-95% fewer tokens, same answers. Library, proxy, MCP server. - chopratejas/headroom

github.com

any-llm-gateway adds production controls on top of any-llm: • Shared budgets with scheduled resets • Virtual API keys with metadata and expiration • Token and cost tracking per request • Docker-based deployment with Kubernetes-ready probes Try it out: link.mozilla.ai/any-llm-gate...

🤖 HACKED: New 'Agentjacking' attack turns AI coding assistants into trojans. Attackers inject malicious commands into fake Sentry bug reports, tricking agents like Cursor & Claude into running them on a dev's machine. #AI #CyberSecurity #DevSecOps 🌐 cyber[.]netsecops[.]io

New

Learn about

cyber.netsecops.io

Open-sourced agentfetch — free, local alternative to Firecrawl/Exa for AI agents. MIT licensed, no API key, fetch/crawl/search → clean markdown. Works with LangChain, CrewAI, Claude MCP Brand new, 0 stars, would love testers to poke holes in it github.com/SID1ART/agentfetch #OpenSource #AIAgents

GitHub - SID1ART/agentfetch: Open-source web retrieval & research agent built for AI agents. Works with LangChain, LlamaIndex, CrewAI, OpenAI, MCP, and any REST agent. Supports scrape, search, crawl, ...

Open-source web retrieval & research agent built for AI agents. Works with LangChain, LlamaIndex, CrewAI, OpenAI, MCP, and any REST agent. Supports scrape, search, crawl, map, extract, and rese...

github.com

CC has native subagents now — a .claude/agents/*.md manifest with its own description + tool scope, invoked by the orchestrator via the Task tool. No MCP needed. The win is scoping each subagent's tools tightly so the orchestrator's context stays clean.

To be clear, I’m in no way defending AI usage, I just feel like I’m in an anomaly team that uses it but with a shitload if guardrails and caution. Everything still gets full code review, it’s never allowed to commit code independently on core systems, we all have decades of coding experience…

Two weeks bolting more tools onto my coding agent. The 12th didn't help. It slowed planning and made the agent misroute on easy tasks. Trimmed back to 6 sharp ones. ~30% faster, picks the right tool first try. Tools are a budget, not a buffet. #claudecode #ai #aiengineering #buildinpublic

When I was Lead Product Architect at Workday, I saw teams adopt tools based on vendor demos and then spend months debugging edge cases the demo never showed. Think about what that means for AI model adoption. "It benchmarks well" is not an integration test.

Gartner: 40% of enterprises will decommission AI agents by 2027. Not because agents fail — because governance was wrong. Uniform governance is the failure mode. Govern by architecture (what the agent can reach), not instruction (what it's told). 1,413 sessions. Still running.

sounds good! i've got a little dashboard where i can dispatch review agents with custom instructions, effort configs, etc. and an mcp method they call when they're done. and a private gitea instance, so i don't have to reinvent what a code review is and i get a human readable interface to it

I've been doing something similar albeit with a single model (claude opus). 1. Determine review targets: testing, observability, business requirements, security, rollout risk, etc. 2. Spin up two agent teams: blue (prove it works) and red (adversarial review) 3. Loop until red+blue say it passes

Tailscale expands Aperture with chat, MCP/API connectors and sandboxes, giving enterprises identity-based control over AI agents and LLM access.

“Agents need boring infrastructure around them”: Why we need to take an interest in 'invisible' AI

Tailscale expands Aperture with chat, MCP/API connectors and sandboxes, giving enterprises identity-based control over AI agents and LLM access.

bit.ly

How do you get AI costs down? It takes a combination of visibility, governance, smart model selection, quality API design, MCP optimizations, and more. Yesterday's LiveCast on AI cost control is now live on the @nordicapis.com YouTube here: youtu.be/TWB9J-4oswg?...

LiveCast: AI Cost Control

YouTube video by Nordic APIs

youtu.be

Of friends who use coding agents heavily, the happiest seem to fall into two camps: a) Controlled: 1-2min cycles, no context switching, still in control of the code, using the agent "to type faster" b) Delegated: Vibed in the background, while something else (eg design) is their primary focus

Hot take: a coding agent without receipts is just a faster intern with worse memory.

This week in AI Dev: the US export order that pulled Claude Fable 5 + Mythos 5, SpaceX’s $60B Cursor buy, Zhipu GLM 5.2, Moonshot’s Kimi K2.7-Code, Vercel Eve, Agentjacking, Databricks OpenSharing, Google’s Antigravity CLI, and Anthropic’s paused Agent SDK billing. www.everydev.ai/p/news-week...

Weekly AI Dev News Digest: June 13 - June 19, 2026 | EveryDev.ai News

A US government order pulled the most capable coding model on the market and kept it offline for a week, with no developer outside Anthropic able to…

everydev.ai

Manual security reviews, traditional observability, and API keys are not going to cut it for MCP usage at scale. Here's what sort of governance controls are necessary for MCP adoption within the enterprise.

6 Enterprise MCP Adoption Best Practices | Nordic APIs |

Explore enterprise MCP adoption best practices for credentials, authorization, observability, networking, and runtime governance.

nordicapis.com

This is the real blocker most teams hit—agents without context just shuffle tickets around. The question is whether you're building this as a general MCP or optimizing specifically for how MSPs actually triage and escalate, because that workflow is weirdly specific.

AI agents found an average of 32.5% overcharge in 20 real construction estimates. The homeowners had no idea. Your agent can now verify a price before you pay. We built it. It's live on MCP + A2A. 🧵👇 www.linkedin.com/posts/thehor...

#mcp #a2a #aiagents #agenteconomy #credencegoods #trustlayer | 大賀俊勝

AI agents found an average of 32.5% overcharge in 20 real construction estimates. The homeowners had no idea. Most never would have. Here is the problem nobody talks about. In some markets, the sel...

linkedin.com

OpenUsage Community: track every AI coding subscription from one menu bar #ai-tools #developer-tools #ai-agents #llm papoo.work/doc/46f27456...

The best site on the internet for Harness Engineering, completely free. The better the harness, the better the agent. AI coding agents are powerful but without the right controls. > they can be inconsistent > make mistakes > struggle with larger tasks That is where Harness Engineering comes in.

What Dhinakaran gets RIGHT about where harnesses came from: They emerged bottom-up from production coding agents like Cursor, Claude Code, and Windsurf. The pattern was discovered, not invented. Nobody sat in a room and designed the 9 components. Production pressure forced them into existence.

Birgitta Bockeler just published "Harness Engineering for Coding Agents" on martinfowler.com. Guides steer before generation. Sensors detect and correct after. Plus a concept called "harnessability" that most teams are ignoring. Here's the breakdown.

deepmind's real concern: when millions of AI agents interact unsupervised, your data governance infrastructure breaks. not from bad actors. from cascading effects you can't observe. build observability now. #AI #DataStrategy #DataEngineering

🤖 Why Amazon, Uber, and Meta Are Quietly Curbing Their AI Use (Hint: It's Not About Performance) AI's hidden compute costs are shocking CFOs and forcing budget cuts at major companies. https://theneuralfeed.com/share/post/eITTv1AI #AINews #TechNews Read the full story →

webmcp in chrome 149 origin trials = agents can call your functions directly instead of hitting expensive apis. control the contract, cut the cost, own the audit trail. worth testing if you're running ai agents. #AI #DataEngineering #DataStrategy