Agents 2.0 is generally available today. The headline change is small to describe: state now survives between steps. It changes what an agent can do in a single run.
From script to worker
Until now, each agent step started in a clean container. Anything the agent installed or wrote was gone by the next step, so runs had to redo work, and long tasks timed out before they finished.
What persists
- The filesystem, for 24 hours
- Running processes, such as a dev server or a database
- Installed packages and caches
Turn it on per agent with one flag:
lattis agents run support-triage --persist 24hWhat stays the same
Isolation does not change. Each run still gets its own sandbox, its own network policy and its own scoped secrets, and nothing is shared between runs.
“State between steps turns agents from scripts into workers.”
Replays
Every run can now be replayed step by step from the console, with the inputs and outputs of each tool call. Replays are free on every plan and kept for 30 days on Team.
