Five companies agreed on a spec this week. That alone is news.
AWS, Cursor, Microsoft, OpenAI, and Vercel published Agent Plugins 1.0 on August 6. Google joined as Core Maintainer on day one. Package your agent skills and MCP servers into a single directory with a plugin.json at the root. Build once. Deploy to ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code.
The "build for one harness" era is over.
What's in Issue #25 of The Agentic Engineer
The Big One: Agent Plugins 1.0
The spec packages Agent Skills + MCP servers into a portable directory. One build, six harnesses at launch. No central registry required — distribute as npm packages, git repos, or zip files. Built on MCP and Agent Skills, both already under the Linux Foundation Agentic AI Foundation. AWS is a founding TSC member.
The skeptic's read: MCP launched with big multi-vendor momentum too and the space is still fragmented. The difference this time is day-one install base. If ChatGPT, Codex, Cursor, Copilot, Kiro, and VS Code all support it at launch, there's already a reason to build for the spec.
Quick Hits this week:
- 🐳 Docker Sandboxes — one microVM per agent, zero host risk.
brew install docker/tap/sbx. #1 on HN at launch (203 pts, 131 comments). Works with Claude Code, Codex, Gemini CLI, Kiro, OpenCode out of the box. - 🚀 Kiro Crew (MIT) — Amazon's internal multi-agent harness. 39,000+ Amazon builders, 500 contributors, 597 updates. Cross-session agents, self-learning, OS-level sandboxing. Zero migration from .kiro config.
- ⚡ YC QM (MIT) — Y Combinator open-sources the harness it runs accounting/legal/engineering on. 11K+ stars. Harness-agnostic: Pi, OpenCode, Codex, Claude Code all work.
- 🧠 Prime Agent (MIT) — one tool (persistent IPython kernel), 95.5% on ARC-AGI-3. Above human expert baseline. Sub-agents as function calls:
await rlm("sub-task"). - 🗄️ DynamoDB Vector Search GA — single-digit ms latency at 99%+ recall for trillions of vectors. Semantic retrieval over agent memory, no separate vector DB to manage.
Tool of the Week: AWS Dogwood
Cedar handles point-in-time policy decisions. Dogwood handles sequences. Write rules like "deny InvokeModel if the agent has already called it 10 times this session." Loop detection and privilege escalation prevention at the infrastructure layer. Apache 2.0, built into AgentCore Policy, zero migration from existing Cedar policies.
Paper Breakdown: The Illusion of Visual Tool-Use
6 multimodal models, 5 benchmarks. Two failure modes found:
- Calling Without Looking: tool call made, evidence received, answer unchanged
- Looking Without Planning: keeps zooming after it already has what it needs
Same pattern as the E3 paper (91% of coding agent tokens wasted). Agents over-call across modalities.
Hot Take: Claude Code Auto Mode Is the Right Call
Users approved 97% of permission prompts. A 97% approval rate isn't a safety gate — it's a latency tax. The honest concern isn't auto mode itself. It's what this normalization enables next. Worth watching, not panicking over.
Read the full issue: https://theagenticengineer.waltsoft.net
The Agentic Engineer is a free weekly newsletter tracking the agentic AI revolution. I read the repos so you don't have to.
Top comments (0)