[Open-Source] I need your worst edge cases to stress-test GenOS, my new AI agent orchestrator.

Hey everyone, I’m currently working on GenOS, an open-source framework for multi-agent LLM orchestration. Under the hood, it uses isolated Rust execution environments and relies on Git worktrees for clean state management and secure sandboxing. The core engine is running smoothly, but before pushing it further, I need to expose it to the harsh reality of real-world use cases. We all know that AI agents (whether single or in swarms) look amazing in demos, but often trip over their own feet the second you take them out of "Hello World" territory. That’s where you come in: what are the real, testable problems you run into when building or using AI agents? I’m looking for concrete, reproducible scenarios to see how GenOS handles them (or if it fails miserably, which will help me iterate). What I'm specifically looking for: Infinite loops & derailments: Tasks where the agent starts hallucinating code execution and just won't stop. State & context management: Swarm scenarios where Agent A forgets to pass crucial info to Agent B, or completely overwrites its work. Isolation issues: Cases where an agent corrupts its workspace by modifying or deleting the wrong files. Complex multi-step tasks: Long workflows where the agent eventually loses track of its initial objective. Drop your use cases, your biggest frustrations with existing frameworks (like LangChain, AutoGen, CrewAI, etc.), or even specific prompts that consistently break your setups. I’ll take the most interesting cases, code them into GenOS to see if the Rust/Git architecture offers a cleaner solution, and I'll report back with the results! Thanks in advance for the feedback You can check it here PISSARAW/GenOS: Git-like branching, deterministic replay, and evidence-driven evaluation for reproducible AI agents.