Blog · Launch

Introducing 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.

Companies are giving AI agents access to repositories, databases, payment APIs and customer records. Sooner or later one of those agents does something consequential, and someone asks three questions: what did it know, what was it allowed to do, and why did it act?

Today the honest answer is usually a log file, written by the same system that's being asked about. It can say that a call happened. It rarely says what the agent had read just before, what it could have read and didn't, which rule let the call through, or what a human approver was actually shown. And because the system under audit wrote it, an auditor has no reason to trust it.

Where this came from

I ran into this building PiPilot, an AI coding platform whose agents work inside people's real GitHub, Vercel and Supabase accounts. When an agent did something unexpected, I had logs but no proof. I couldn't show what it had seen, whether it had been allowed to do what it did at that moment, or why a check had passed. Every company putting agents into production is heading for the same wall, and auditors, regulators and customers will be the ones asking.

What Decision Horizon does

Decision Horizon sits on the agent's tool path. That can be in front of an MCP server, in front of an OpenAI-compatible model API, or around a function in Python or TypeScript. Before every consequential action it writes a signed Agent Decision Record, and after the action it writes a second record with the outcome. The intent record holds:

  • who sponsored the action, and which agent took it;
  • the authority it held at that instant, including what it was explicitly not granted;
  • what it had read, and what it could have read but didn't;
  • which policy rule decided, and what a human approver was shown;
  • what the record itself cannot see.

Records are Ed25519-signed and hash-chained. Anyone with the public key can verify a chain offline, and the verifier catches reordering, truncation, splicing, a forged record under the right key id, and an action that started but never reported an outcome.

The same layer enforces policy. In the demo, a prompt injection tells an agent to copy a secret into a public folder. The write is blocked. The next day, in a new session, the agent tries again and is blocked again, and the record names where the secret was first seen.

What's shipped

  • decision-horizon 0.9.2 on PyPI: the CLI, the MCP proxy (dh proxy), the OpenAI-compatible gateway (dh gateway), the Python SDK and the verifier.
  • @decision-horizon/sdk 0.5.0 on npm, for Node agents. Its records verify with the same dh verify.
  • A public verifier at verify.decisionhorizon.dev.
  • Export to SIEMs in OCSF and CEF.

Before launch, the design went through ten rounds of structured adversarial review. Each of the 124 findings was either fixed or published as a stated limit. That isn't an accredited audit, and we don't claim it is: there's no SOC 2 and no third-party penetration test yet.

What we're not claiming

A record is tamper-evident, not tamper-proof: an alteration is detected, not made impossible. A call that never passes through Decision Horizon leaves no record, so every record names its insertion point and lists what that point can't see. Where the record can't know something, it says unknown rather than guessing. The honesty section on the home page and the limits in the docs spell this out.

Try it

pip install decision-horizon
dh keygen
dh init
dh proxy -c dh.toml -- npx -y @modelcontextprotocol/server-filesystem /path/to/repo

The quickstart walks through recording and verifying a first session. If you run agents with real permissions and want to see one of your own workflows recorded end to end, book a time with me.

Want to see a record of one of your own agent workflows? We'll replay one with you.

Book a demo
What an Agent Decision Record contains →