Investors deck · v0.1 · updated 2026-07-10 — press ⌘P / Ctrl+P to save as PDF.
Meet Pete.
He builds agentic systems for other companies — support agents, ops agents, finance agents. He is good at it. The models work.
The hard part was never the model. It is the part where the agent touches money.
Every client draws the same line.
The agent can answer tickets, tag, summarize, draft. The moment it reaches for an action that moves money, the client pulls it back.
“Agentic systems can’t be reliable. I’m not putting one near the money — I can’t risk the loss.”— Pete’s client, every time
So a human still sits in the seat.
The workaround is always the same: keep a person in the loop to approve every money-moving action by hand.
It works. It also means the agent never quite does the job it was hired for. The human became the product.
But the line is already moving.
Some teams have crossed it. ElevenLabs ships support agents that handle account inquiries, transaction disputes, and collections for retail and financial-services customers — routine cases resolved without a human in the loop.
Source: elevenlabs.io — AI agents for financial services. The frontier isn’t should agents move money. It’s who can prove what they did.
Observability is funded. Control is missing.
Coralogix just raised $200M to scale AI-native observability. That validates the urgency: enterprises need infrastructure for agents operating production systems.
But watching an action after it runs is not control. By the time a dashboard shows a duplicate payment, the money is already gone.
KIFF sits between the agent’s decision and the consequential action. It refuses what should not run, then produces a receipt whose trust boundary lives outside the customer’s own database. Observability tells you what happened. KIFF decides what is allowed to happen.
A normal audit log is still a database row someone with access can change.
KIFF doesn’t reach inside the agent.
We don’t touch the model, the prompts, or the reasoning. We sit on the one boundary that matters — between the agent and the consequential action — and check it against the current state of the thing it touches, before it runs.
action
A harness around the action, not a rewrite of the agent. Five lines against the code Pete already has.
So Pete gets more than a log. He gets a record he can’t be lied to.
Every decision becomes a signed record — what was proposed, the state it was checked against, the decision, and the policy behind it. Signed, tamper-evident, verifiable. Pete can hand it to an auditor without trusting his own database. The record is a protocol output — the proof the boundary ran — not the thing being sold.
state = FULLY_REFUNDEDAnd the alerts that make Pete look good.
KIFF doesn’t just record the save — it tells Pete the moment it happens. He walks into the room with the evidence already in hand.
PAID · 9 retries stoppedFULLY_REFUNDED · $150 not paid outPROMISE_TO_PAY · FDCPA / CONC windowSame agent. Same speed. Now with a record that holds and a signal Pete can act on.
The thesis: the boundary that lets you cross the line.
Pete’s story is one shape of a general problem. Every agentic system rebuilds the same machinery — state, valid actions, authority, approvals, evidence — around a new agent each time. KIFF makes that a reusable domain you build once and connect any agent to, without rewiring the system underneath. Software is moving from code that does what it was written to do, to agents that decide at runtime — so the foundation has to own a boundary that can refuse a consequential action before it runs — with a record that holds outside the agent’s own reasoning. Cross the line with peace of mind.
KIFF normalizes the mechanics of that domain — and the refusal at its boundary — the way TCP/IP normalized byte transport:
Not vertical-specific. Any system whose actions map to a state machine can carry an external, verifiable layer of truth.
Pete crosses the line — and we make it the default.
The moat isn’t observability — that only says what happened after. It’s the reusable operational domain teams build on KIFF — state, valid actions, authority, approvals, evidence — that any agent connects to and is replaced against without rewiring, and that every governed operation makes costlier to leave.
See it live: the agent proposes, KIFF decides against the current state, and every decision lands in the dashboard as a signed, tamper-evident receipt you can open and verify.
We’re raising to turn this proven wedge into a category. What we’re raising, what it buys, and the milestones it unlocks — live. This deck is a starting point, not a data room.
kiff.dev · kiff.dev/whitepaper · proofs: /proof