Use krino

Shadow, report, enforce

The safe path to a cheaper agent: record decisions in shadow mode, read the report, then enforce tool selection when the data agrees.

krino never asks you to trust it up front. Every decision starts in shadow mode, the report shows you the evidence, and you turn on enforce mode only when that evidence is good.

Run in shadow mode

Shadow is the default for every decision. krino asks the decision model in the background, adds no wait to your agent's steps, and changes nothing. It writes one JSON line per step to a local trace folder.

TypeScript

import { createKrino } from "@krinolabs/krino";
 
const krino = createKrino({
  projectName: "my-agent",
  decisionModes: { toolSelection: "shadow", riskGate: "shadow" },
});

Run your agent on its usual work. In a short-lived process, call await krino.flushAll(DEFAULT_FLUSH_TIMEOUT_IN_MILLISECONDS) before you exit, as the quick starts do. Decisions still open after that wait are recorded as cutOff.

Read the report

Terminal

npx krino report

The report shows one block per decision and mode, for example Tool selection · shadow:

Line	What it tells you
Calls	How many decisions krino asked for, by status.
Agreement	How often krino's tool set covered what your agent actually used. Per step on the AI SDK, per run on the Claude Agent SDK.
Saved if enforced	The estimated saving, with cache reads, cache writes, and the decisions' own cost included.
Added latency	What your agent waited. Always 0 ms in shadow mode.

Below the blocks are cache health (for all runs and for multi-step runs), the cut-off count, and one "next step" line.

Check the bar for enforce

The report's next-step line suggests enforce mode for tool selection only when all of these hold:

At least 20 compared samples per host.
Agreement of 90% or more on every host it judged.
A positive estimated net saving.

Before that, look at two health checks. More than 5% of decisions cut off means your process exits before krino can answer: flush before exit. A cache read share below 50% on multi-step runs means prompt caching is probably off or broken.

Enforce tool selection

Set tool selection to enforce. Keep the risk gate in shadow: in v0.1, createKrino rejects riskGate: "enforce".

TypeScript

import { createKrino } from "@krinolabs/krino";
 
const krino = createKrino({
  projectName: "my-agent",
  decisionModes: { toolSelection: "enforce", riskGate: "shadow" },
});

In enforce mode, krino waits up to decisionTimeoutInMilliseconds (800 ms by default) and then sends only the selected tools, once, at step 0. If the decision model is slow, fails, or is less confident than minimumConfidence (0.8), your agent gets all its tools.

Keep watching

Keep running npx krino report. In enforce mode, 5% of runs (explorationRate) skip enforcement and send all tools, so agreement stays measurable. Enforced runs see only the suggested tools, so the report compares exploration runs only.