Engineering the Lean Agent: Inside Agav’s War on Autonomous Token Bloat Agav says its approach to autonomous coding agents — treating agentic execution as a systems scheduling problem with prefix cache stabilization and isolated subagent forks — cuts inference costs in half while still handling complex repository orchestration. The company frames the problem as unmanaged context accumulation, circular repair spirals and naive prompt expansion that exhaust developer budgets. Autonomous coding agents routinely exhaust developer budgets through unmanaged context accumulation, circular repair spirals, and naive prompt expansion. Agav reframes agentic execution as a systems scheduling problem, deploying prefix cache stabilization and isolated subagent forks to slash inference costs in half without abandoning complex repository orchestration.