What you build on Pragmatic, and how it runs today.
Long-running agents survive process restarts, deploys, and infrastructure failures, then resume exactly where they stopped. A run that dies nine hundred steps in does not re-bill nine hundred model calls; it picks up at step nine hundred and one, with every tool executed exactly once.
Reproduce the exact run that failed, step by step, including the nondeterministic model behavior that caused it. Bugs you could never reproduce become runs you can replay.
Every run has a replayable, tamper-evident record of what the agent actually did: every model turn, every tool call, and what each cost in tokens, signed when a runtime key is configured. When an auditor asks what the agent did, the answer is the Journal.
Self-hosted, today. The runtime links into your agent as a library, and your Journals stay on your own disk, in your own environment. A managed cloud backend with a hosted replay console is in development.