June 16, 2026 · 7 items (site) · 9 items (base)
Anthropic now lets Managed Agents execute tools (Bash, files, code) inside a customer-controlled container, behind their firewall. Outbound-only connections — Anthropic never initiates inbound. Private MCP servers are now supported. This is the missing piece for regulated sectors: health, finance, legal.
Agent SDK and non-interactive claude -p now pull from a separate monthly credit: $20 (Pro), $100 (Max 5x), $200 (Max 20x). Unused credit does not roll over. A structural shift for teams building on Claude.
Microsoft Agent Framework 1.0 GA (merged AutoGen + Semantic Kernel). CrewAI: 52.4k stars, 2 billion agent runs in 12 months. Google ADK in 4 languages. MCP surpasses 200 server implementations. ACP merges into A2A under Linux Foundation. 8 major frameworks in active competition.
ServiceNow-AI published an extended benchmark for evaluating AI agents: 3 domains, 121 tools, 213 scenarios. Measures tool selection, multi-step reasoning, error recovery (failed tools, unexpected results), and resource efficiency. Fills a major gap in agent evaluation.
H Company published an agent that controls GUIs entirely on consumer hardware — no cloud needed. Keyboard/mouse automation, screen interaction, app control. Open weights, variants 0.8B to 35B on Hugging Face. A privacy-first alternative to cloud offerings.
IBM argues production success depends on robust agent logic, not just the underlying model. Four pillars: multi-step reasoning with fallback, reliable external system interaction, long-term state management, graceful error handling. Teams should invest in agent architecture, not chase benchmarks.
DevArt tested Gemma 4 12B with Ollama + OpenCode on real dev tasks: landing page, bug fixes, UI generation, mini-game — all 100% local, zero API keys. The creator admitted he was wrong: this local stack actually works for production development. A credible privacy-first alternative to cloud agent coding.
Watch for adoption rates in finance and healthcare. If the self-hosted sandbox clears compliance hurdles, it could unlock enterprise agent deployment at scale.
If MiniMax publishes the M3 weights as promised, it's the first open-weight model to match closed-source frontier on SWE-Bench Pro (59%). A seismic shift for open-source agentic development.
The battle is shifting from "best model" to "best agent ecosystem." Self-hosted infrastructure, dedicated billing, framework consolidation, and agent-specific benchmarks are all maturing in the same window. The agent stack is becoming a product category.