Start-of-Session Smoke Test

Every session of a long-running agent should orient itself and run an init script plus a basic end-to-end check before starting new work, so a broken inherited state is caught before anything is built on it.

Evidence: moderate

An agent that resumes long-running work inherits whatever the previous session left behind, and among the failures Anthropic observed was a session leaving the environment broken. If the next session starts building straight away, it stacks new work on a fault it never noticed and later struggles to tell old damage from new.

Anthropic's answer is a fixed start-of-session ritual. The agent runs pwd, reads the git log and the progress file, picks the highest-priority unfinished feature, starts the development server through init.sh, and runs a basic end-to-end check before doing anything new. The point of the check is to catch a broken state early, while it is still cheap to attribute. The same post reports that agents often marked features done without verifying them end to end, and that explicitly prompting them to test through browser automation, as a user would, markedly improved results. That makes the smoke test the entry-side counterpart of verify before done, and the two together bracket one unit of work per session.

The evidence is one lab engineering post describing its own harness, so it is a well-motivated practice rather than a measured one.

The ritual generalises beyond coding. An agent that operates live systems should begin every run, and every resumption after a crash, with a read-only baseline of the systems it is responsible for before it diagnoses or acts. This matters most after an interruption, where the agent cannot assume the world matches its records; see unknown state on resurrection. The orientation half of the ritual depends on externalised progress artefacts being there to read.


Sources

  • Anthropic Engineering (Justin Young), 26 Nov 2025, 'Getting up to speed' — https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents