Benchmarking Lightweight AI-Agent Frameworks on a 2 GB Raspberry Pi 4: Framework Overhead, Offline Feasibility, and Fleet Density

A new class of "lightweight" AI-agent frameworks marketed for edge deployment appeared in 2026, yet no clean, apples-to-apples measurements exist for how they actually behave on constrained hardware. We benchmark five such frameworks (ZeroClaw, PicoClaw, NanoBot, smolagents, and Agent Zero) on a single dedicated Raspberry Pi 4 with 2 GB of RAM, holding the underlying language model constant so that the differences we measure reflect framework overhead (memory footprint, agent-loop latency, out-of-the-box capability) rather than inference cost. Four frameworks ran and produced 100 trials across a five-task suite (five trials each); the fifth (Agent Zero) could not be installed or run on the device and is reported as a documented negative result. Three findings stand out. First, on a hosted (cloud) model all four runnable frameworks reach high task success, but their memory footprints separate cleanly by runtime: the compiled agents are 3 to 6 times lighter (ZeroClaw, Rust, ~20 MB; PicoClaw, Go, ~31 MB) than the interpreted ones (smolagents, Python, ~90 MB; NanoBot, Python, ~130 MB). Second, on a fully offline (local) model no agent completed even the simplest task: across eight agent x model combinations the two feasible local models fail for opposite reasons, a capable model (Qwen2.5-1.5B) being too slow for the frameworks' default network timeouts and a fast model (SmolLM2-360M) lacking the tool-calling the frameworks require. The model runs offline; the agents do not. Third, as a positive counterpoint, a single 2 GB Pi hosts eight concurrent instances of every framework at 100% success with linear RAM growth, and the compiled frameworks pack roughly 10 to 20 times denser than the Python ones. We release the harness, raw data, and analysis to make the comparison reproducible.