The words we use

A program that runs tool calls on your behalf: Claude Code, Codex, a LangChain app. A registered agent is one you told Krynix about; an undeclared agent is one seen in the evidence that nobody registered.

One run of an agent from start to finish, kept as a chain of records where each record carries the hash of the one before it.

One record in a session: a prompt, a tool call, its result, or a decision.

The agent doing something: running a command, editing a file, calling an API.

Krynix’s answer to “may this tool call run?”, taken from the policy in force for the agent at that moment: allow, or deny with a reason.

Also called: in the evidence API, /v1/decide; a call no decision covered is “unchecked”.

The rules that decide an agent’s tool calls. A policy is bound to an agent; the organisation’s baselines apply to every agent as well, and the strictest answer wins.

A policy the organisation applies to every agent, beside each agent’s own.

Which of two things Krynix is doing for your agents. Monitoring: it records what they do. Enforcement: it also decides their tool calls. Krynix infers the posture from the evidence, never from what was declared.

Also called: monitoring posture, enforcement posture.

What you said is wired up for an agent: that Krynix decides its tool calls. The inventory compares it with what the evidence shows.

An agent declared enforced whose sessions show tool calls no decision covered. The evidence contradicts the declaration.

A standing assertion about every run, such as “no run deletes production data”. Krynix evaluates each control against each session when the session closes. A control is not a check that runs once.

One control against one session. The verdict is pass, fail, or unknown when the evidence to decide it was not there. Unknown is never a pass.

A discrepancy the independent verifier found in the evidence: a record missing or altered, a decision not made by the policy in force, a call made while the agent was suspended.

The chained records of a session, together with what signs them.

The signature over a session’s chain of records. It proves nobody altered the records since they were signed.

The portable export of one session’s evidence, for an auditor or a court.

Proof that a session was deleted: when, why, and what its chain looked like.

A session that retention never sweeps, whatever the retention window says.

Every tool call of this agent is refused, whatever its policy says. It stops only an agent that asks Krynix before each call.

What an install does while Krynix is down. Block tool calls: nothing runs unchecked, and the agent waits. Keep agents working: the agent carries on, and the calls made meanwhile are not checked.

Also called: in the evidence API, fail closed and fail open.

Which parts of a run an install records: prompts, tool arguments, tool output. A part an install does not capture is “not captured”, never “nothing happened”.

The credential an install carries. It can only add evidence; it cannot read the record. Revoking it disconnects every machine using that install.