ActTruth by SOVERYN

ActTruth is an honesty and accountability layer for AI agents that take real actions — calling tools, spending money, writing to systems. The problem it targets is specific and familiar to anyone running agents in production: agents rarely announce their failures. They retry silently, half-complete a task, report success, or loop while quietly burning API credits and vendor spend. ActTruth's answer is an act ledger. Every action an agent takes is recorded in a verifiable log, including the ones that went wrong, so a developer reviewing a run sees what actually happened rather than the agent's own summary of it. The second control is an unprompted spend allowance: a hard budget for actions the agent initiates without an explicit human instruction. That distinction matters more than a flat rate limit, because the expensive incidents are almost never the calls a human asked for — they are the retries, the self-directed research loops and the tool chains an agent decided to run on its own. Capping that class of spend separately gives teams a way to let agents act autonomously without accepting an open-ended bill. It is Apache-2.0 licensed and supports local receipt export, so the ledger is inspectable and portable rather than locked inside a vendor dashboard — a reasonable requirement if you are using it as an audit trail for spend or for post-incident review. The roadmap points at a Lattice-class memory layer, which would extend the same accountability idea from single actions into agent memory. Honest limits: this is early infrastructure aimed at developers, not a no-code monitoring product. You need to instrument your agent for the ledger to be useful, and it complements rather than replaces general observability tooling. If you are running agents that spend money or mutate production systems, though, the combination of a failure-inclusive ledger and a separate autonomous spend cap addresses a gap most agent frameworks leave open.