Tandem hardens autonomous agent runtime

This is the kind of infrastructure work that separates a demo agent from something you can actually trust in production.

• –Three-state outcomes (`passed`, `needs_repair`, `blocked`) make retries budget-aware instead of binary.
• –Repair briefs turn telemetry into actionable context, which is much more effective than replaying the same prompt.
• –Classifying tool outputs by usefulness, not just execution, is the right move for web search and read-heavy workflows.
• –Cross-checking self-report against telemetry catches the most common agent failure: plausible excuses that contradict the log.
• –The unresolved frontier is semantic quality: detecting when output is thin, but not obviously broken.