kiff.dev First, watch a real AI agent work an end-to-end task while KIFF decides each consequential action before it runs. Then step through the product itself.
An agent doing actual work, mitigating an incident, moving on production, with every risky action allowed, held for a human, or blocked against the system's live state.
You want an agent that can mitigate an incident, roll back, scale, fail over, without doing something irreversible under pressure.
In this clip The on-call agent rolls back a bad deploy (allowed), a region failover is held for a human to approve, and "drop the database" is refused outright, each decided against the service's live state, every one on the record.
The moment an agent is wired to real systems it can trigger consequential actions; you need to see the blast radius and know the wrong move is stopped.
In this clip A scan surfaces every action the agent's code can reach, then live: the safe fix runs, a high-blast-radius action waits for a human, and a schema migration mid-incident is blocked, the move that has wiped live databases this year.
Short walkthroughs of KIFF Cloud itself: author a governed domain, watch the control room decide, and read the signed evidence.
You have an agent making tool calls today and want it governed, without a rewrite.
In this clip From the Runtimes page: mint a key, drop in the language-agnostic call, and the agent shows up as a connected, governed runtime, with the open-source adapters on GitHub.
Turn refund limits, approval thresholds and state rules, today scattered across services, into one versioned contract your agents decide against.
In this clip Using the kiff-domains skill so your coding agent writes the kiff.yaml, then validating and activating a real refunds domain in the dashboard.
When an agent acts on production, you need to see what it proposed and how each call was decided, not hope it behaved.
In this clip The Govern view: decision counts by outcome and a state-aware decisions table where a refusal reads as decided against the entity's real state, ending on a signed receipt.
Weeks after an incident, compliance asks exactly what happened and what stopped it. "Trust me" isn't an answer.
In this clip The per-domain activity trail and a signed, tamper-evident receipt: the action, the state it was checked against, and the verdict, recorded before anything ran.
Connect an agent, govern one consequential action, and see the decision in minutes.