New tool Cartha tracks what your AI agents are actually doing
Running one AI agent is manageable, but once you have several running at once, it becomes hard to see what each agent remembers, what it's calling, or what it's costing — until something breaks and you're stuck reconstructing events from logs afterward. Cartha was built to solve that. It's SDK-first: three lines of Python with a decorator on your agent function (TypeScript support coming next).
It offers three main features. First, trace replay lets you click into any run and see the full reasoning chain — what memory was pulled, what tools were called, the actual decision path — not just raw logs. Second, scoped memory enforces access at the user/agent/team/org level, so if a support agent shouldn't see a finance agent's memory, it genuinely can't, rather than just being told not to.
Third, breaks spending down per agent and per tool call instead of one lump sum per run, which is typically where teams discover where the real waste is happening. It's -, natively supports MCP and A2A at the SDK level, and is . The builder is now looking for people who actually build and operate agent systems to try it and give honest feedback on rough spots in the developer experience, places the abstraction breaks down, or problems it might be solving that don't actually exist.
Key points
- Applied with a 3-line Python decorator; TypeScript support planned next
- Trace replay shows the full run: memory pulled, tools called, decision path
- Memory access is enforced by scope (user/agent/team/org), not just policy-based
- Cost is broken down per agent and per tool call to surface where money is actually going
- -, natively supports MCP/A2A, and is