Bottom line：Perplexity introduced Teammates, persistent autonomous agents designed to handle continuous background workflows and collaborative tasks in enterprise workspaces.
Functionality：Automates multi-step workflows like deep research, data analysis, ticket triaging, and code review across isolated environments.
Access：Rolling out on the web for Max tier subscribers, with desktop and mobile applications forthcoming.
Bottom line：Cursor released a direct app publishing feature that enables users to build, secure, and deploy functional web applications directly from the editor.
Features：Automatically provisions databases, runs deterministic and agentic security checks, and generates live shareable deployment links.
Availability：Available immediately to all paid tier subscribers via the latest editor update.
U.S. Midterm Elections and Local Data Center ResistanceMMultiple Sources
Bottom line：Following midterm elections and local ballot measures, tech markets faced sharp pullbacks amid growing regulatory scrutiny over data center expansion.
Policy Impact：Voters across multiple states overwhelmingly approved municipal ballot measures to restrict or slow down power-heavy data center developments.
Market Reaction：Major indices saw significant single-day corrections as investors re-priced infrastructure capital expenditure risks.
Bottom line：Meta opened public access to Muse AI real-time video avatars across the United States, offering expressive interactive voice and video chat.
Technology：Powered by Muse Realtime Voice and Avatar models, synchronizing dialogue with natural expressions and gestures.
Customization：Users can interact with customizable avatars capable of handling interrupted conversations dynamically.
Recursa: Recursive Self-Improvement via Skill RepositoriesIIndependent AI Researchers
Bottom line：Researchers proposed Recursa, a self-improvement framework where models construct validated skill libraries to substantially boost task success without modifying base weights.
Methodology：Parallel learners propose task fixes and workflow instructions, which are verified and merged into a structured domain repository.
Performance Gains：Raised DeepSeek-V4.1-Flash success rates from 30.0% to 55.0% on held-out legal benchmarks using only self-generated guidance.
Collaborative Intelligence BenchmarkSStanford, MIT, and Oxford
Bottom line：Researchers introduced a collaborative intelligence benchmark to evaluate how frontier language models function as cooperative teammates under operational stress.
Evaluation Scope：Tests models across eight collaborative dimensions including conflict resolution, error reporting, and team coordination.
Key Findings：While models perform well in normal settings, performance drops significantly when team members introduce simulated behavioral hurdles.
ServerBench: Automated LLM Serving Stack TuningCCarnegie Mellon University
Bottom line：CMU researchers demonstrated that coding agents can autonomously design and optimize complete LLM serving architectures to outperform baseline efficiency.
Evaluation Metric：Scored systems based on latency-qualified token value per dollar of GPU operational cost.
Results：Agent-optimized serving designs delivered up to 4.87x the value per GPU dollar compared to reference implementations.
Bottom line：JetBrains launched its 2026.3 IDE lineup integrating ‘JetBrains Air’, a dedicated layer for running, tracking, and verifying multi-agent coding workflows.
Integration：Supports major coding agents like Claude Agent, Codex, Gemini, and GitHub Copilot directly inside IDE tabs.
Review Tools：Provides side-by-side diff views and editor intelligence to help developers review agent modifications safely.
