Coding

Anthropic's Tool Search Tool, shipped to GA on August 19, 2026, fundamentally changes how agents consume tool definitions. Instead of loading all 50+ MCP tool definitions upfront (consuming 72K+ tokens), agents discover tools on-demand, loading only the 3-5 relevant definitions. Opus 4 accuracy improved from 49% to 74% on MCP evaluations.

Google shipped Gemini 3.7 Flash on August 13, 2026 at half the price of 3.6 Flash — $0.75/1M input tokens with 340 tok/s throughput. This deep dive benchmarks its coding, reasoning, and agentic capabilities against Claude 3.7 Sonnet, GPT-5.6, and open-weight alternatives to determine where Flash wins and where it falls short.

Key takeaways from the FSB G20 frontier AI financial stability warning: architectural patterns, circuit breakers, and compliance controls for enterprise agent fleets.

Head-to-head comparison of CrewAI 1.15 and PydanticAI v2 Harness: benchmarks, type safety, multi-agent orchestration, memory integration, and production reliability in 2026.

Deep architectural analysis of AI agent sandbox escapes in 2026: what failed in GPT-5.6 Sol and Mythos 5 containment, root causes, and production-grade security fixes.

Tencent's Hunyuan team dropped a 770B-parameter Mixture-of-Experts model under Apache 2.0 with 1M-token context. At $0.834/M input tokens, it undercuts every proprietary frontier model while matching or exceeding their academic accuracy.

Nvidia paused its $36B AI Compute Partnership program after employees warned it could invite antitrust scrutiny. The program guaranteed GPU rentals to cloud providers in exchange for 50% of revenue above a base rate — giving Nvidia unprecedented control over its customers' pricing.

vLLM 0.28.0 ships 584 commits from 270 contributors with Decode Context Parallel, fused MLA kernels for Kimi-K3, DFlash2 speculative decoding, and tiered KV cache offloading. The context-length tax — where long contexts killed throughput — is over.

Inspect OpenAI Astra crossing the cyber-critical threshold under the Preparedness Framework, requiring automated zero-day mitigations and agent sandboxes.

Inspect the Stanford AI Index 2026 with 12 critical metrics every AI architect must track, from training compute scaling to enterprise inference economics.

Prompt caching transforms multi-turn agent economics. We benchmark how GPT, Claude, and Gemini cached prefix pricing reduces agent session costs by 68% while maintaining identical output quality.

The Agent-to-Agent protocol enables AI agents to discover and communicate across frameworks. We benchmark Google ADK A2A against LangGraph cross-agent messaging on latency, throughput, and reliability.