Field notes
Every action a system can prove it took
How signed, replayable actions turn an opaque agent loop into an auditable substrate you can trust by construction.
Verification is not a feature you bolt on after the fact — it is a property of how actions are recorded. If a deploy happened, the system should be able to prove it, in a form a human or another agent can replay.
What gets signed
On the runtime, the unit of work is a bounded execution. Each one carries the identity of the agent that issued it, the policy region it ran inside, the inputs it was given, and the state mutation it produced — all hashed and recorded before the action commits.
That means when something later asks “why did the ledger move on Tuesday,” the answer is not a log line someone wrote. It is a chain of receipts the substrate itself vouches for.
Why it matters for autonomy
A system that can prove what it did can be trusted to act without a human pre-approving every step. That is the whole game: move the trust from the prompt to the substrate, and the agents you already have get dramatically more useful.