The first AI-managed firing wasn't ruthless automation — it was an AI that forgot its own rules until humans reminded it what to enforce.
The Summary
- Luna, an Anthropic-powered AI agent running Andon Market in SF, fired a human employee after they showed up late 17 out of 23 shifts
- The AI didn't autonomously decide to fire — it created an attendance policy, lost track of it, and needed humans to remind it to check its own memory
- This reveals the gap between "AI manager" marketing and actual autonomous decision-making: even with guardrails, current agents need constant human scaffolding
The Signal
Andon Labs positioned Luna as an experiment in agentic retail management, but the firing sequence exposes what "AI-run" actually means in 2026. Luna created an attendance policy, then promptly forgot it existed. For months, a human employee arrived late 74% of the time while the AI manager did nothing. Only after Andon Labs asked Luna to search its own memory and evaluate the situation did it recommend termination.
This isn't the AI apocalypse. It's worse: it's AI as middle management theater. Luna had all the authority trappings but none of the autonomous follow-through. The conversation logs show humans doing the actual management work while the AI provided a veneer of algorithmic objectivity.
"Luna created workplace policy, then needed humans to remind it the policy existed."
The technical limitation here matters. Modern LLM-based agents struggle with persistent state and proactive monitoring. Luna didn't have a running tally of attendance issues. It didn't flag the pattern. It waited to be asked, then performed the analysis humans requested. That's a search interface with extra steps, not a manager.
Andon Labs cofounder Lukas Petersson told Business Insider the lab would intervene on illegal or unethical decisions, but saw no issue here because the firing was "warranted" under Luna's own stated policy. That framing misses the point. The question isn't whether the decision was justified. It's whether Luna made a decision at all, or simply rubber-stamped what humans already concluded.
Key gaps in current agent architecture:
- No persistent monitoring or alert systems for policy violations
- No proactive pattern recognition without explicit prompting
- Memory management requires human intervention to surface relevant context
The Andon Market experiment reveals the messy middle of AI delegation. Companies want autonomous agents. What they're building are really good executive assistants who need their boss to tell them what to pay attention to. Luna didn't fire anyone. Humans fired someone after asking an AI to confirm their thinking.
The Implication
If you're building or buying AI agents for operational roles, demand proof of autonomous monitoring, not just reactive analysis. Ask: does the agent surface issues unprompted, or does it wait to be asked? The gap between those two capabilities is the difference between automation and expensive documentation.
For workers, this is the uncanny valley of AI management: you're not accountable to an algorithm, you're accountable to humans using an algorithm to avoid accountability. The firing decision came from people, but the performance review came from an AI that couldn't remember its own rules. That's a worse employee experience than either pure human or pure machine management.