Notes on agent evidence.
What we're building, why, and what we've learned about proving what an AI agent knew, was allowed, and did.
Decision Horizon for Claude Code, Cursor, Copilot and every coding agent
Coding agents act through built-in tools that no MCP proxy sees. One command now puts a policy decision, an Ostiary score and a signed record on every one of their tool calls.
ProductOstiary: a doorkeeper for every agent action
Rules catch what you thought to write down. Ostiary learns how each agent normally behaves, scores every tool call before it runs, and signs its reasons into the record.
ProductAuthorization decides. Evidence proves.
The agent-security field has converged on checking every tool call before it runs. That stops the bad action. Proving later why an action was allowed takes something more.
Field notesCapture at decision time
Decision provenance, bitemporal data, and why reconstructing an agent's decision after the fact goes wrong. The most trustworthy evidence is the record written at the moment.
Field notesThe hosted record store and the Console
Your records can now live in a hosted store you never have to trust for the truth, and your security team can search and verify them in a browser.
ProductWhat it would take to make the record lie
Tamper-evident is not tamper-proof. Here is what an attacker would need for each kind of lie, and what every record says about itself.
Trust modelWhat an Agent Decision Record contains
Eight blocks and one signature: who sponsored the action, what authority it held, what the agent had read, which rule let it through, and what the record cannot see.
The recordIntroducing Decision Horizon
When an AI agent does something consequential, nobody can say why it was allowed. We built a signed record of what it knew, was allowed, and did, at the moment it acted.
Launch