In the early days of AI agent deployment, everyone was obsessed with latency and token costs. But as we move deeper into 2026, the most dangerous bugs aren't the ones that crash; they're the ones that succeed quietly and return nothing. Dev.to user Tommybuilds, who builds AI agents for small businesses, recently documented a scenario where a lead-finding agent ran for 13 rounds, returned 8 qualified leads, and threw zero errors. Yet, five of those rounds returned absolutely nothing because the agent was being bot-blocked and silently moving on.
The Illusion of Success
The core problem isn't just the empty results; it's the metrics we use to judge them. Tommybuilds noted that if he had only looked at the final count of "8 leads on day one," the agent would have looked like a massive success. The run log, however, revealed that five consecutive rounds hit sites that flagged the agent as a bot and served empty pages. The agent did exactly what it was told: search, find nothing, move on. No exception, no warning, no retry storm. It was a perfect execution of a flawed instruction set. This pattern extends beyond lead generation. In another instance, a status panel for model endpoints showed all green while the underlying accounts had been logged out for two days. The health check was verifying that the process was alive, not that a request could actually be served. Similarly, a session cookie reported a validity of over a year, but the server killed it in under three days. The system was measuring something adjacent to the work, not the work itself.
Stale Rules and Silent Zombies
Long-running agents rely on standing rulebooks to maintain consistency. Tommybuilds enforces a hard cap of 140 lines for these indexes, forcing developers to delete or merge old rules. But when a retired rule wasn't fully removed, a fresh session obeyed it and stopped dead on a sign-up page. The agent wasn't broken; it was following orders nobody meant anymore. This is the digital equivalent of a zombie process, consuming resources and blocking progress while appearing fully operational to the dashboard.
The True Cost of Completion
As we optimize for reliability, we also need to rethink how we calculate value. Price per token is becoming irrelevant compared to cost per finished task. Tommybuilds cited data from the SemiAnalysis GDPx board, noting that while DeepSeek Flash might cost $0.94 per task with a 5.4% completion rate, and Opus 5.5 costs $51.41 with a 12% rate, the real cost per *finished* task is roughly $17 versus $428. If your workflow can tolerate failures and retries, the cheaper model wins. If it can't, the expensive one might actually be more cost-effective. The metric that matters is not what you spend, but what you actually get.
Key Takeaways
- Empty results must be logged as results, not skipped, to catch silent bot-blocking or parsing errors.
- Health checks must hit the real endpoint or page, not just verify that the process is running.
- Session validity should be verified by opening the actual settings page periodically, not trusting cookie expiry dates.
- Agent rulebooks need hard size caps and mandatory deletion of retired rules to prevent 'zombie' behavior.
- Calculate cost per finished task, not per call, to determine the true efficiency of your model selection.
The Bottom Line
Your agent isn't broken; it's just obedient. In 2026, the most expensive failure mode is silence. If your dashboard is green but your outputs are empty, you aren't running an AI agentβyou're running a very expensive, very polite placeholder.