The Agent Runtime Wars
In a single three-week window this April, every major AI provider — Anthropic, OpenAI via AWS, Microsoft, and Google — shipped a hosted agent runtime. The convergence is no accident. A year of enterprise pilots stalled on the same three rocks — durable state, sandbox isolation, observability — and the industry has settled on a single answer: stop selling tokens, start selling agent-hours.
The harness and sandbox layer that frontier labs perfected on their own coding agents is now a procurable line item, IAM and audit trails included. Two camps are forming. Model labs want to host the agent themselves and keep the margin close to the model. Hyperscalers want a neutral runtime where any model can run, and the compute beneath is theirs. It is the cloud-era value-capture battle, replayed one layer up.
By 2027, running an agent may be as commoditized as running a container — and the buyer’s question will shift from which model to use, to which runtime to bet on.