HarnessRL co-evolves training harness with policy for open-ended RL

AWS's post walks through five production challenges in multi-agent systems and how Deep Insight addressed them, noting agents can be stood up in days with open source frameworks and cloud services but production is where the real challenges begin.

Qwen Code shipped v0.26.0 alongside a Desktop build, adding session-targeted Skill and Extension refresh in the daemon and warm session-create for Workspace-isolated bound Sessions via the runtime broker. The release also bounds initial publication guard waits in the managed-agent path.

Blackstone, with $1.3 trillion in assets under management, says it is the largest owner of compute outside the hyperscalers, backing neoclouds and data centers. Its BXN1 unit committed $5 billion to Crux AI, a neocloud partnered with Google to commercialize TPUs, and also backed chip startup Etched.

Alter gives each agent its own scoped identity and keys instead of one shared over-scoped credential, with default-deny policy checks on every request. One SDK call handles storage, refresh, policy and audit for 140 providers, with LangChain, MCP and AWS integrations.

The release adds /claude-api api-version to move Claude API code to version 2026-10-21, plus a claude_code.safety_event OpenTelemetry event fired whenever a permission rule, auto mode, safety check, hook or sandbox stops or prompts a tool call.

Mammoth-1 (27B) ranks #1 on the Jev Decision Index; Elephant-1 (9B) is tied for #1 on the Artificial Analysis Decision Index. Both take a 16k-token context window and return choices, probabilities, and scores for classifying, routing, and scoring text.

V7 Go builds a Context Graph from company files that agents query for deal screening and insurance underwriting. On V7's hardest graph-query tests across thousands of documents, GPT-6 Astra reached 89% accuracy.