How are you handling silent failures in multi-step agent workflows?

Working on multi-step agent systems (agents calling tools, other agents, APIs), I keep running into a frustrating class of failures:

One step returns slightly malformed or incomplete state Downstream steps continue executing anyway The issue only surfaces several steps later

Nothing actually “fails” — the system just produces the wrong result.

These are hard to catch:

Logs/traces help explain what happened after the fact But they don’t prevent bad execution from propagating

I’m experimenting with:

Explicit state validation between steps Blocking unsafe transitions Replay from intermediate failure points

Curious how others are handling this in production. Are you relying purely on tracing/logs, or enforcing stricter contracts between steps?

I’m building something in this space and looking for a few design partners to try it out (happy to wire it up myself): https://www.agentsentinelai.com/