Descriptions:
Craig Hewitt breaks down four major updates Anthropic announced during a 3-hour livestream focused on the Claude multi-agent ecosystem, alongside a surprising infrastructure partnership that took effect the same day. The biggest technical update enables agents to spawn sub-agents on demand for parallel task execution, task specialization (e.g., a dedicated security review agent), or model escalation — spinning up Claude Opus 4.7 for high-stakes subtasks while running the main orchestration on Sonnet, with support for up to 25 concurrent threads reporting back to the main context.
Three additional updates round out the release: outcome-based agent instructions that let developers define success criteria in plain English and have the agent work backwards to achieve them; webhooks for multi-agent sessions that eliminate polling on long-running jobs; and official documentation for “dreaming,” an asynchronous memory consolidation process that prunes and refines up to 100 sessions into a lean output memory store to prevent context bloat in long-running agents.
On the infrastructure side, Anthropic announced a partnership with SpaceX granting full access to the Colossus 1 supercomputer — over 300 megawatts of additional capacity going live within the month. The immediate effect was a doubling of Claude Code’s 5-hour rate limits for Pro, Max, and Team plans, removal of peak-hour throttling, and substantially higher Opus API rate limits. Hewitt contextualizes this alongside Anthropic’s broader compute acquisition strategy and the developer frustration that built up over months of opaque quota management.
📺 Source: Craig Hewitt · Published May 07, 2026
🏷️ Format: News Analysis
