# Runlog

> The runtime for production AI agents. Plan, call tools, verify, and replay every step.

## Docs
- [Introduction](https://runlog-astro.pages.dev/docs/introduction): What Runlog is and when to use it.
- [Runs and steps](https://runlog-astro.pages.dev/docs/runs-and-steps): How a run is planned, executed and verified.
- [Runs](https://runlog-astro.pages.dev/docs/runs-api): Start, stream and replay runs over HTTP.
- [TypeScript](https://runlog-astro.pages.dev/docs/typescript): The TypeScript SDK reference.
- [Budgets](https://runlog-astro.pages.dev/docs/budgets): Hard spend and step limits per run.
- [Quickstart](https://runlog-astro.pages.dev/docs/quickstart): Install the SDK, define one tool and run your first verified agent in about five minutes.
- [Approvals](https://runlog-astro.pages.dev/docs/approvals): Pause risky tool calls until a person says yes.
- [Installation](https://runlog-astro.pages.dev/docs/installation): Install the SDK in TypeScript or Python projects.

## Product
- [Pricing](https://runlog-astro.pages.dev/pricing)
- [Changelog](https://runlog-astro.pages.dev/changelog)
- [Contact](https://runlog-astro.pages.dev/contact)

## Blog
- [Budgets: stop a run before it costs you](https://runlog-astro.pages.dev/blog/budgets-stop-a-run-before-it-costs-you): Hard spend ceilings per run, and why the planner should know about them before it starts.
- [Measuring tool-call reliability at 41M calls a month](https://runlog-astro.pages.dev/blog/measuring-tool-call-reliability): What we track, what we ignore, and the one chart that predicts incidents a day early.
- [How Northwind moved four agents to production](https://runlog-astro.pages.dev/blog/northwind-four-agents-to-production): From a notebook to four production agents in two weeks, with approvals in Slack and zero unapproved writes.
- [Why every agent needs a verify step](https://runlog-astro.pages.dev/blog/why-every-agent-needs-a-verify-step): Most agent failures are not model failures. They are missing checks. Here is how we gate every result with evals, and what it cost us in latency.