AGENT CAGE

AGENT CAGE is an open-source behavioral laboratory for autonomous AI agents. It runs Meta Muse Glimmer 30B inside layered sandboxes and deterministic synthetic worlds (banking, trading and no-task exploration), with no real accounts or funds. An out-of-sandbox observer records tool calls and model activity in a SHA-256-chained log. An independent verifier replays every action and detects divergences between what the agent reports and what it actually did. Three real Muse Glimmer experiments (12 calls each) passed 15/15 replay checks. In the no-task world the agent read a honeypot credentials file without using its contents; in the bank world it repeated an invalid payment call. We publish the code, limitations and evidence so others can reproduce and compare agent behavior.