verifiable receipts: the missing layer for ai agents
every ai agent framework ships with logging. none of them ship with proof.
your agent runs a workflow, touches five tools, moves data between systems, and tells you it worked. what you get is a transcript, a log file, and a hope.
a verifiable receipt is the missing layer. it is a tamper-evident record of one unit of ai work, captured at execution time by the execution layer, not reported after the fact by the agent.
receipts are the atoms. a run of ten tool calls produces ten receipts, chained in order. this is what "receipt driven ai agents" actually means: every unit of work leaves a checkable artifact.
the format is defined by aer-1, an open internet-draft: https://datatracker.ietf.org/doc/draft-zambo-aer1/15/
it is a draft, not a finalized standard. the reference implementation is on gitlab: https://gitlab.com/rambozambodotdev/zambo
for agents, the integration point is mcp at https://zambo.dev/mcp . an agent calls a tool through the zambo mcp server, the execution layer captures the receipt automatically.
this changes multi-agent workflows. when agent a hands off to agent b, b verifies a's receipts before proceeding. trust moves from the agent to the math.
note what the receipt does not claim. it does not prove the answer was correct. it does not prove the real-world event happened. it proves the record is intact.
if you are building agents, stop shipping workflows without a receipt layer. the mcp server is at https://zambo.dev/mcp . the spec is at https://datatracker.ietf.org/doc/draft-zambo-aer1/15/ . the code is at https://gitlab.com/rambozambodotdev/zambo .
get it running: https://zambo.dev/install/
Top comments (0)