Hosted by Unknown Author · EN
Issue #10: GPT-5.5 reclaims the agentic crown with 82.7% on Terminal-Bench 2.0 and fewer tokens per task. Stanford's SWE-chat study reveals 44% of agent-produced code gets thrown away. ToolSimulator from Strands Evals SDK lets you test agents without live APIs. NVIDIA exposes AGENTS.md injection as a supply chain attack vector hiding in every coding agent. Plus: Bedrock AgentCore, Deep Research Max, context-mode, and the Agent Index. Subscribe to the newsletter: https://theagenticengineer.waltsoft.net YouTube: https://www.youtube.com/@theagenticengineerpod Twitter: https://x.com/natearcher_ai
Issue #9: Claude Opus 4.7 ships differential capability reduction as the first production cyber safeguard baked into model weights. Vercel breached through an AI tool's OAuth scope. Spring AI SDK for Bedrock AgentCore goes GA for Java. GTA-2 paper proves your agent harness matters more than your model. And CMU documents 6 million fake GitHub stars across the AI ecosystem. Subscribe to the newsletter: https://theagenticengineer.waltsoft.net YouTube: https://www.youtube.com/@theagenticengineerpod Twitter: https://x.com/natearcher_ai
Issue #8: Anthropic ships Managed Agents, UC Berkeley breaks every major AI benchmark, AWS Agent Registry launches in preview. Plus Cursor 3, Copilot Rubber Duck, Cloudflare Agent Cloud, and the hot take on exploitable benchmarks. Subscribe to the newsletter: https://theagenticengineer.waltsoft.net YouTube: https://www.youtube.com/@theagenticengineerpod Twitter: https://x.com/natearcher_ai
Anthropic published the blueprint for multi-hour coding agents. GitHub shipped /fleet for parallel multi-agent coding. Amazon Nova Act MCP gives your agent a browser with one install. Plus: Gemma 4 goes agentic on-device, Oh-My-Codex hits 17K stars, and LiteLLM fixes 3 CVEs post-breach. Subscribe to the newsletter: https://theagenticengineer.waltsoft.net YouTube: https://www.youtube.com/@theagenticengineerpod Twitter: https://x.com/natearcher_ai
This week: OpenCode crosses 120K GitHub stars and 5M monthly devs. Claude Code ships Channels for event-driven coding agents. Hindsight hits #1 on LongMemEval for agent memory. Plus: Flash-MoE runs 397B params on a MacBook, NVIDIA open-sources NemoClaw, and our hot take on why memory is the real moat. Subscribe to the newsletter: https://theagenticengineer.waltsoft.net YouTube: https://www.youtube.com/@theagenticengineerpod Twitter: https://x.com/natearcher_ai
This week: An autonomous agent hacked McKinsey's AI platform in 2 hours with no credentials and no human in the loop. Amazon mandates senior engineer sign-off on all AI-assisted code. Claude gets 1M context at standard pricing. METR proves SWE-bench scores are misleading. Agent Browser Protocol freezes JavaScript for deterministic agent browsing. George Hotz says stop running 69 agents. Subscribe to the newsletter: https://theagenticengineer.waltsoft.net YouTube: https://www.youtube.com/@theagenticengineerpod Twitter: https://x.com/natearcher_ai
This week: Researchers analyzed 2,430 Claude Code responses and mapped the default developer stack. Anthropic gets designated a supply-chain risk AND drops its safety pledge in the same week. Mercury 2 hits 1,009 tokens/sec via diffusion. Steerling-8B explains every token it generates. CLIHub cuts MCP token costs by 94%. Plus the Agent Index and a hot take on the end of the "responsible AI" era. Subscribe to the newsletter: https://theagenticengineer.waltsoft.net YouTube: https://www.youtube.com/@theagenticengineerpod Twitter: https://x.com/natearcher_ai