Post

📼 Introducing tapes: transparent AI agentic telemetry. In every single AI session or workflow, there's an incredible amount of lost information on the "what", "why", and "how". www.youtube.com/watch?v=ATeU...

As @bizza.pizza would put it: tapes "is long term care for your agent sessions". This is how we build the agentic platforms of the future with durable audit trails, memory, secure workflows, infrastructure, and strong loopback capabilities.

How's it work? Tapes has a number of components that enable it to have huge scale: * a storage driver (sqlite locally by default) * an LLM provider (ollama locally by default) * an embedding provider (ollama locally by default) * a vector store driver (sqlite-vec by default)

$ tapes serve brings up 1) the proxy and 2) the api server. The proxy connects to your configured LLM provider and transparently captures events and messages. The API server makes commands like "search" and "checkout" work

$ tapes search "query" lets you search back on previous conversation threads and sessions: since we use content addressing, you'll see the hash for that point in the conversation turn.

$ tapes checkout <hash> lets you do context check-pointing and retry. Using a hash from the session graph, you'll be able to then use "$ tapes chat" to time travel to a point-in-conversation and retry it.

Next up on our roadmap is partnering with Cursor to help tackle Agent-Trace: tracking the metrics and agent stats from AI coding assistance tools in our storage layer and search capabilities. agent-trace.dev