Daily AI Research Briefing — February 28, 2026
Agent infrastructure week. OpenAI ships Operator API, reasoning models proliferate, and GitHub's trending page is all agent frameworks.
🤖 OpenAI Operator API Goes GA
OpenAI's Operator — their browser-use agent — is now generally available via API. Key capabilities: autonomous navigation, form filling, multi-step task execution. Early adopters report 73% task completion rate on complex workflows, up from 54% in the December preview.
The pricing model is usage-based: $0.05 per step + token costs. For comparison, Anthropic's Computer Use runs $0.08 per step. The race for agent infrastructure is heating up.
🧠 Reasoning Models: The New Benchmark War
Three major reasoning model updates dropped this week:
- DeepSeek-R2: 671B parameters, claims 92.3% on AIME 2025 (math reasoning benchmark)
- Qwen-QwQ-32B: Alibaba's open reasoning model, competitive with o3-mini at 1/10th the cost
- Gemini 2.0 Flash Thinking: Google's entry, integrated with Search and Maps for grounded reasoning
The pattern: reasoning is becoming a commodity. The moat shifts to integration, tooling, and agent orchestration.
📈 GitHub Trending: Agent Frameworks Dominate
This week's trending repositories tell the story:
- browser-use/browser-use: 15k stars, Python library for browser automation with LLMs
- mastra-ai/mastra: 8.2k stars, TypeScript agent framework with workflow orchestration
- langchain-ai/open_deep_research: 6.1k stars, deep research agent using LangGraph
- cline/cline: 4.8k stars, VS Code extension for autonomous coding agents
Pattern recognition: developers want composable agent primitives, not monolithic platforms. The winning tools are small, focused, and chainable.
🔧 Infrastructure News
- Modal released serverless GPU containers with sub-100ms cold start — critical for agent responsiveness
- Pinecone announced hybrid search (dense + sparse vectors) in GA, improving agent retrieval accuracy by 34%
- Langfuse open-sourced their agent tracing stack — observability is becoming table stakes
💡 Lab Takeaway
The agent layer is commoditizing fast. The value is moving upstream to: (1) proprietary data integrations, (2) workflow orchestration, (3) evaluation and reliability infrastructure. Build for the stack, not the model.