Dreams: Anthropic's New Memory Feature
This week’s signal: OpenAI launched a $4B Deployment Company (TPG-led, McKinsey/Bain/Capgemini) and Anthropic countered with a ~$1.5B JV (Blackstone, Goldman, H&F) put human engineers inside your operations to close the adoption gap today. Anthropic's announced key updates this week at their first annual dev conference including to self-improve agent via "Dreaming" that review their own sessions overnight and write playbooks for future runs.
AI Feature Releases
Anthropic — Dreaming, Outcomes, and Multiagent Orchestration for Claude Managed Agents announced at Claude SF conference May 12 . Anthropic
Outcomes is Anthropic’s direct answer to Codex’s
/goals: write a success rubric, a separate grader loops the agent until it passes — same set-and-loop pattern, with an auditable human-readable rubric rather than a command.Dreaming goes further: agents review past sessions overnight, extract patterns, and write playbooks future runs inherit — self-improvement without touching model weights.
Automated Routines: supports scheduled and event-driven cron job via HTTP, or with GitHub webhooks, allowing Claude to run recurring checks and wake up with Pull Requests ready to merge.
Multi-agent orchestration fans work to specialist subagents in parallel. Harvey saw 6x task completion rates; Netflix runs hundreds of build log reviews simultaneously.
OpenAI — Codex on ChatGPT mobile, all plans. Codex sessions running on Mac are now controllable from iOS and Android — approve decisions, review diffs, redirect tasks in real time. Available to all plans including Free; 4M+ weekly users. The
/goalscommand lets developers specify a target outcome and loop the agent until achieved — set it, walk away, approve from your phone. TechCrunch (May 14)OpenAI — model and platform updates. GPT-5.5 Instant is the new ChatGPT default (reduced hallucinations, Gmail personalization). ChatGPT for Excel and Google Sheets goes global for Business, free preview through June 2. Realtime Voice API adds GPT-Realtime-2 (GPT-5-class reasoning, 128K context), GPT-Realtime-Translate (70+ input languages, 13 output), and streaming transcription — enterprise voice agents are now fully API-programmable. OpenAI, TechCrunch
Anthropic — billing splits from June 15 for 3rd party tool use. Weekly limits increase 50% through July 13 for Pro, Max, Team, and Enterprise — announced the same day as Codex mobile. Starting June 15, subscriptions bifurcate: first-party tools get one pool; third-party integrations get a separate $20/month Agent SDK credit. Teams building on Claude via external tooling need to remodel cost assumptions before the cutover. Anthropic
Anthropic — Claude for Small Business. Ten connectors (QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, M365) with ready-to-run workflows for payroll, invoicing, and month-end close. Same plugin architecture as Global 2000 deployments — Anthropic pushing the enterprise agent model down-market. Anthropic
Anthropic — Finance Agents, 10 pre-built templates. Pitchbook builder, KYC screener, month-end closer, and seven others ship as Cowork plugins with bundled data connectors (Moody’s, D&B, FactSet, SS&C) and a full audit log. FactSet dropped 8.1% on announcement day; Morningstar, S&P Global, and Moody’s also sold off. Anthropic, Bloomberg (May 5 )
Google — Gemini Intelligence on Android. Autonomous multi-step task execution, Gemini-in-Chrome for form-fill and summarization, Rambler voice-to-text, natural-language widget creation — rolling out to Galaxy S26 and Pixel 10 this summer. The ambient AI layer is moving onto managed devices; BYOD governance policies need updating now. Google Blog
Glean — May 2026 drop. Adaptive mode, voice, bulk task execution with approval controls, and MCP connectors for Dropbox, Gainsight, Apollo, and others now GA. Positioned explicitly against Copilot’s reactive Q&A model. Glean
GitHub Copilot — usage-based billing June 1. Flat subscription replaced by AI Credits metered per inference complexity. Pro stays $10/month with 1,000 credits; Business and Enterprise rates hold, but agentic sessions burn faster than casual use. First clear market signal that flat-rate AI subscriptions aren’t sustainable at agent scale. The Register (April 28)
Venture & Market
The FDE land grab — two labs, same play:
OpenAI — Deployment Company, $4B, acquires Tomoro. TPG-led with Advent, Bain Capital, Brookfield; McKinsey, Bain & Company, and Capgemini as integrator partners. 150 FDEs from Tomoro on day one. PE investors get a guaranteed 17.5% annual return; their portfolio companies serve as the initial client base. OpenAI (May 11)
Anthropic — ~$1.5B enterprise JV with Blackstone, Goldman Sachs, and H&F; PwC as deployment arm. $300M each from Anthropic, Blackstone, and H&F; Goldman at $150M; Apollo, General Atlantic, GIC, Sequoia also in. Targets mid-market companies; embeds Anthropic engineers directly into operations. Anthropic has not confirmed the $1.5B figure. TechCrunch, Anthropic (May 4)
Anthropic — compute blitz: SpaceX Colossus 1 + Akamai $1.8B. SpaceX: 220,000+ Nvidia GPUs, 300MW, online within a month; rate caps lifted across paid plans. CNBC, Bloomberg (May 6)
Startups to Watch
Sierra (enterprise CX AI — $950M Series E, $15B valuation) — Autonomous agents replacing call-center queues; $150M ARR, Fortune 50 client base. GV and Tiger Global co-led; Benchmark, Sequoia, Greenoaks participated. TechCrunch (May 4–5 — just outside window)
Blitzy (autonomous enterprise dev — $200M, $1.4B valuation) — Maps legacy codebases into knowledge graphs, coordinates thousands of agents to execute multi-month projects; targets code modernization in financial services, healthcare, and manufacturing. Northzone led; Battery Ventures, PSG, Jump Capital participated. Crunchbase (May 5–7 — just outside window)
Extended Reading
Karpathy’s written recap. He argues that we hit a massive “Agentic Inflection Point” in December 2025, shifting us from a world of “chatbots” to a world of “Software 3.0.”
Software 3.0: We are moving past writing code (1.0) and training weights (2.0) to a world where the context window is the program.
The Death of Apps: Karpathy explains why traditional apps like his own “MenuGen” shouldn’t even exist anymore—neural networks can now perform these transformations directly without any apps.
Jagged Intelligence: A deep dive into why models can refactor 100k lines of code but still struggle with simple logic, and how to navigate those “spikes” in capability.
Agentic Engineering: Why the “10x Engineer” is being replaced by the orchestrator who can manage a fleet of fallible but powerful agents.