SpaceX Deal Unlocks Reliable Agent Scaling

Anthropic secured exclusive access to SpaceX's Colossus supercluster, addressing compute bottlenecks from surging Claude Code demand. This doubled subscription rate limits, eliminated Pro/Max peak-hour caps, and raised API limits up to 17x for top tiers. Builders gain consistent performance for agentic coding, eliminating frustrations like OpenClaw restrictions and slow Opus 4.7 inference. Result: Shift platforms from text endpoints to fully hosted models with harnesses and unlimited scaling, letting you deploy without infra headaches.

Managed Agents Features Deliver Production Wins

Deploy Claude Managed Agents in an afternoon for serverless execution. Key features:

  • Memory: Store expertise as markdown folders (global for editorial rules, personal for prefs like em-dashes). Pulls only relevant files per request, speeding responses without bloating prompts. Spiral uses it to apply style guides automatically.
  • Multi-agent orchestration: Coordinator (Haiku 4.5) spins Opus 4.6 Fast subagents in parallel. Spiral's multi-draft requests dropped from serial 20-30s delays to parallel, cutting costs by a third via cheaper models. Use when parallelism or model mixing pays off; skip otherwise to avoid coordination overhead and debug complexity.
  • Outcomes: Grader AI loops writer against dynamic rubric (global standards + user memory). Spiral deploys soon to enforce writing quality.

Mitigate lock-in: Log runs to your DB for data portability; build custom tools on your servers (any model inside). Trade-off: Agents tied to Claude, but tools escape vendor limits.

Dreaming Automates Institutional Learning

Dreaming (research preview) analyzes up to 100 past sessions/memory, merging duplicates, resolving contradictions, and extracting patterns into cleaner stores. Builds 'compound engineering'—each run improves the next via collective team knowledge, not per-user repeats. Spiral tests show early gains; extend to Claude Code for repo-specific tastes without messy manual memory. Outperforms static files by self-organizing, trading minor overhead for quality.

Platform Insights: Harnesses > Models, Infra > Prompts

Generic model-agnostic harnesses fail—Anthropic tests show model-tuned ones yield 'drastically different' results, making swaps secondary to optimization. Infrastructure walls (sandboxing, uptime, storage) block most builders; Managed Agents handles it, freeing focus. Agents stale quickly—assign owners or build meta-agents for self-upgrades. Anthropic's 'outcome + budget' philosophy plus auto-subagent selection points to self-managing fleets.