If capacity constraints are behind recent Claude issues, we should expect service improvements as new compute comes online. Worth diversifying to other providers until reliability returns to previous levels.

This $200M commitment spans four years across global health, education, and economic mobility. For context, that's roughly equivalent to Anthropic's Series C funding round, suggesting significant organisational focus on beneficial deployments rather than just scaling revenue.

The state machine approach addresses a real production problem: agents with 40+ tools often fail to make progress. By constraining tool access per phase, this makes smaller models more effective than throwing more compute at the problem.

For production agent systems with variable execution paths, this platform could solve scaling bottlenecks we'll hit as agent workloads grow. Worth tracking as these non-deterministic patterns become more common.

If we're building conversational interfaces, this approach reduces the grunt work of training utterance variations. The LLM-backed NLU handles ambiguous requests better than traditional rule-based systems.

The two-stage tokenization approach (continuous OHLCV data → discrete tokens → transformer training) could extend beyond finance to any time-series domain. This pattern may become standard for specialised foundation models targeting structured sequential data.