AI Agents Approach AGI: Sequoia's 2026 Prediction and the Race Toward Autonomous Systems

Sequoia Capital recently published an analysis titled "2026: This Is AGI" examining the rapid advancement of long-horizon AI agents. The firm observes that these agents now demonstrate the ability to take actions and iterate over extended periods, with exponential progress occurring in this research domain. Sequoia estimates that by 2028, agents will reliably complete tasks at a level comparable to human experts, positioning this capability as a litmus test for Artificial General Intelligence.

Cursor's research on scaling long-running autonomous coding reveals that GPT-5.2 models outperform Opus 4.5 and GPT-5.1-codex for extended autonomous work. The analysis emphasizes that model selection becomes critical for extremely long-running tasks, and that many improvements emerge from reducing complexity rather than adding it. The research concludes that while infrastructure and models matter, prompt engineering remains the most significant factor.

Personal Intelligence and Contextual AI

Google introduced Personal Intelligence for Gemini, a beta feature that connects the AI assistant to Gmail, Photos, YouTube, and Search to deliver personalized responses based on user data. The system employs "Context Packing" technology to extract and compress only relevant information rather than processing entire context windows. Available initially to Google AI Pro and Ultra subscribers, the feature remains disabled by default for privacy protection.

Google is also testing Gemini Auto Browse for Chrome, an agent-style feature that would enable AI to browse the web, manage tabs, and interact with Chrome autonomously. Code analysis suggests this may launch as a premium Gemini Ultra feature.

Research Challenges Conventional Prompting

A Google Research paper challenges established prompting practices by demonstrating that simply repeating prompts twice improves accuracy across Gemini, GPT-4o, Claude, and DeepSeek. The technique won 47 out of 70 benchmark tests with zero losses, with some tasks showing accuracy improvements up to 76 percentage points. The method works because large language models process text left-to-right, and repetition allows tokens to reference the full query for additional context without increasing latency or output length.

Major AI Companies Approach Public Markets

OpenAI, Anthropic, and SpaceX have begun preliminary work on potential IPOs, with combined valuations approaching $2 trillion. SpaceX is valued at $800 billion, OpenAI around $500 billion, and Anthropic near $350 billion in recent funding discussions. If all three proceed, they could exceed the total capital raised from approximately 200 U.S. IPOs in the previous year.

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, lost two co-founders who returned to OpenAI. Co-founder and CTO Barret Zoph departed along with Luke Metz and Sam Schoenholz. The startup had raised $2 billion at a $12 billion valuation less than a year prior.

Infrastructure and Compute Partnerships

OpenAI signed a multi-year agreement with Cerebras worth over $10 billion to receive 750 megawatts of compute through 2028, aiming to boost inference speed across OpenAI's products. The partnership integrates Cerebras' wafer-scale systems into OpenAI's inference stack to reduce latency for real-time AI responses.

Character.ai doubled production inference speed through GPU workload tuning and hardware-level optimizations, reducing both latency and operational costs across their systems.

Healthcare AI Applications

OpenAI released ChatGPT Health, integrating with Apple Health, wearables, and electronic health records to analyze longitudinal patient data trends rather than simply responding to symptom queries.

Anthropic launched Claude for Healthcare, targeting clinical workflows including medical coding, prior authorizations, and patient history summarization. The model cites medical literature sources like PubMed for verification.

Google DeepMind released MedGemma 1.5, an open-weight model that interprets 3D medical images including volumetric CT scans and MRIs, expanding beyond 2D X-ray analysis.

Model Releases and Technical Advances

GPT-5.2-Codex became available in the Responses API, an upgraded version optimized for agentic coding tasks. It supports four levels of reasoning effort settings, has a 400,000-token context window, and costs $1.75 per million input tokens and $14.00 per million output tokens.

Claude Code introduced MCP Tool Search, allowing Claude Code to dynamically load tools into context. When MCP tool descriptions would consume more than 10% of context, tools are loaded via search instead of being preloaded.

Ministral 3 launched as a new family of dense language models with 3B, 8B, and 14B parameter variants optimized for low-resource environments. The models support image understanding and were trained using Cascade Distillation, an iterative distillation and pruning method.

Zhipu AI, creator of GLM language models, launched GLM-Image, their first open-source image model.

Financial Services AI Integration

Affirm updated its underwriting system to incorporate real-time signals including account balances and cash flow trends, enabling more informed credit decisions at checkout.

FIS launched a platform in partnership with Visa and Mastercard that enables banks to securely support AI-initiated payments within existing card network frameworks, with availability expected by Q1 2026. The platform aims to authorize agent-driven transactions and enhance fraud protection as agentic commerce scales.

Moneyhub will deploy its AI-powered transaction categorization and enrichment engine across Nationwide's 16 million customers, analyzing payments and adding context including merchant identification, location data, and payment details to help manage spending and detect fraud.

Enterprise and Consumer Tools

OpenAI rolled out ChatGPT Translate, a translation tool offering language-specific translation, fluency improvements, tone adjustments for business and academic contexts, and text simplification. The feature is free to use and does not require a paid account.

Google revamped its Trends Explore page with Gemini AI, adding features that surface related search terms, auto-generate comparisons, and suggest follow-up queries.

Slack released an integrated AI agent to function as a personal assistant for searching within channels.

Perplexity partnered with BlueMatrix to enable searches of equity reports with direct chat responses.

SimilarWeb partnered with Manus, a Meta-owned AI agent platform, to integrate web traffic and engagement data directly into the chatbot. The integration provides 12 months of domain history, pageviews, users, sources, and segmentation data for marketing analysis.

Commerce and Standards

Shopify and Google co-developed the Universal Commerce Protocol (UCP), an open standard for AI agents that defines discovery and negotiation mechanisms between agent and merchant.

Airbnb hired Ahmad Al-Dahle, Meta's former AI leader who led generative AI and the Llama models team, as its new CTO. The move signals Airbnb's push toward building an AI-powered travel concierge for search and personalized trip planning.

Global AI Adoption

Microsoft research shows global AI adoption reached only 15.1% by mid-2025, with 1.2 percentage point growth. Highest adoption rates were observed in UAE (64%), Singapore (60.9%), and Norway (46.4%). The United States ranks 24th (28.3%), while China ranks 61st (16.3%). South Korea showed the largest growth at 4.8 percentage points.

Strategic Decisions and Market Positioning

Anthropic blocked third parties from using the Claude Code API, a decision that drew criticism as it may push users to other model providers rather than converting them back to Claude Code.

Elon Musk stated that the new Grok 4.20 will excel at various functions but will not surpass Claude, acknowledging that Anthropic "did something special" with their model development.

Want more AI updates?

Visit https://bosq.dev/blog for more posts like this, plus practical guides and curated links. If you enjoyed this roundup, share it with someone on your team.

References:

Tags: #AGI #AIAgents #SequoiaCapital #OpenAI #Anthropic #GoogleGemini #PersonalIntelligence #AIInfrastructure #HealthcareAI #AIAdoption #MachineLearning #LargeLanguageModels #AICompute #AIPlatforms #EnterpriseAI