What does an agentic system know when its agents disagree?
Memory belongs to the agent. Operational reality belongs to the system.
Models can reason and tools can execute, but neither gives an operation one shared answer to what happened, what is true, what may happen next, and who has authority. KIFF makes that reality explicit and executable.
Could operational reality become reusable infrastructure?
The intelligence was new. The plumbing was not.
Each system rebuilt the same foundations: entities, events, state, valid actions, permissions, approvals, execution results, and evidence. The repetition was hiding in the glue around the agent.
Define the operational pieces once.
A domain can package the thing being acted on, what is true about it, what may happen next, and who has authority. Those pieces can be composed instead of recreated for every use case.
Let the agent become replaceable.
When agents propose through a stable operational contract, teams can maintain, upgrade, or replace them without rebuilding the surrounding business system each time.
Do not make the agent carry operational reality.
An agent is good at interpreting a situation and proposing a move. It should not also be the only place that knows the current state, the valid actions, who approves what, and what must be recorded. When all of that lives inside one agent implementation, maintenance becomes a rewrite.
KIFF does not make the agent smarter. It gives every actor one shared, enforceable, and replayable operational reality.
The same pieces can support different agents and different domains.
KIFF gives these pieces a common shape without deciding what they mean for your business. That stable shape is what makes composition possible: a future agent can enter through the same action contract, and another system can read the resulting state and evidence without a new one-off integration.
Not every action needs KIFF. Three conditions define the fit.
Whether the action is allowed depends on the current state of the entity, not just a rule written in a doc.
Money moves, or the action is irreversible.
An auditor, a regulator, or a cyber-insurer asks what was authorized.
KIFF decides every proposed action before your system runs it, and where the action requires issued authority, it decides against the card you gave that agent. What makes both possible is an operational domain: the entity, events, state, valid actions, authority, approvals, and evidence defined once. A card narrows what that domain already allows, which actions this agent may take, on what, and how much, so agents and services can propose work against a reality none of them owns or recreates.
Its vocabulary, business rules, risks, evidence requirements, and definition of success.
How state, actions, authority, approvals, execution outcomes, audit, and replay fit together.
Normalize the mechanics, not the semantics. Reuse the contract, not one agent's implementation.
Reusable infrastructure only helps when the system uses it.
KIFF does not invent your domain's truth, repair missing events, or intercept a path that bypasses it. A card narrows what the domain and its permissions already allow; it cannot make a forbidden action legal, and it is required only for the actions your domain marks as needing one. Your systems must publish reliable facts, and the consequential calls you want governed must pass through the runtime or guard. Inside that path, the operational contract remains stable even as agents change.
Keep your code. Put the contract in front of the consequence.
Already running agents that can move money, change infrastructure, or touch customer data?
A 30-minute review, founder-led: what those agents can reach with nothing on the path that can refuse them, what a card would have to say to bound it, and what evidence exists today if a customer, an auditor, or a regulator asks. You keep the findings whether or not KIFF turns out to be the right answer.
book a 30-minute review →