Agentwall middleware adds human approval layer to stop runaway AI agent tool calls

Developers have released Agentwall, a Python and TypeScript middleware layer designed to intercept AI agent tool calls before they execute, addressing a critical execution-control gap in current agent frameworks. The tool was motivated by real failure modes such as agents issuing production-wide database deletes due to missing filters, or entering infinite retry loops against non-idempotent APIs. Agentwall classifies each tool call as safe, cautious, or destructive using deterministic local rules defined in a configuration file, without relying on a secondary language model for judgment. Destructive calls are held pending human approval, which can be routed through a CLI prompt, Slack workflow, or internal dashboard, and are denied by default if the approval service is unavailable. The middleware is framework-agnostic, supporting OpenAI, Anthropic, and LangChain tool formats, and logs every attempt as a JSON event including blocked calls.