Intencion · Product analytics for AI agents

3 min read Original article ↗

Agent

Intents7d30dAll

What the agent doesRunsSuccess

Generate a part drawing1,24095%

Check a tolerance stack76091%

Detect an interference29054%

Run a stress analysis48086%

Generate a CNC toolpath new90build next →

Flagged this week: 1 intent to fix, 1 to build next.

Works with

Auto-instruments

OpenAIAnthropic

Ships issues to

LinearGitHub

01 What you see

See what your users actually want.

Three views of your traffic: the goals users bring, what your agent did on each run, and the requests it can't handle yet. This is the product layer for your agent. Every run carries an outcome you defined, so a row reads completed or failed at a step, and tells you whether the goal was served.

Refund a charge96%

Track my order92%

Change my plan61%

Cancel subscription79%

verify_user · auth · 38ms

lookup_order · orders-db · 42ms

notify_user · email · 19ms

✓ success

Book a callbacknew118

Split a paymentnew64

Pause my accountnew41

Gift a subscriptionnew27

02 How it works

From a run to a fix, or a feature.

Add the SDK once. After that it's the same loop every week: see what's failing, fix the biggest one, build the request people keep making.

i.

Capture

Patch your client once. Every run is recorded: model, tokens, latency, and a success or failure verdict you define, not a model's guess.

ii.

Label intent

Each run is labeled with the user's goal as a named business intent like refund_request. You declare it, or we infer it per run from the input.

iii.

Group & rank

Runs with the same goal are grouped. Failures and missing features sort by how often they happen.

iv.

Fix & ship

You fix the top problem and build the top request.

v.

Improve

The numbers go up. Next week, you do it again.

03 It compounds

Success rate climbs every week.

Each fix raises one intent's success rate. Over a month, that's the gap between an agent people put up with and one they trust.

Success rate, week by weekillustrative

78%84%89%93%Week 1Week 2Week 3Week 4▲ +15 points in a month

W1fixed identity verification

W2fixed a refund edge case

W3shipped order tracking

W4shipped callback booking

04 Close the loop

Don't just see the work. Ship it.

Turn any failure or unmet request into a Linear or GitHub issue, pre-filled with the evidence, in one click.

failure · 61%

Change my plan

43% of failures: can't verify identity

LinearGitHub

Fix: Change my plan succeeding at 61%, can't verify identity

Impact160 failed runs/wk (~96 tickets)

Wherefails at verify_user (auth), no fallback

☐ success ≥ 85%☐ add OTP fallback

bugagentfrom-intencion

05 Install

Patch your client once.

Every call is captured: model, tokens, latency, and outcome. It works at the class level, so calls your framework makes on its own are caught too. TypeScript and Python.

Read the docs →

agent.ts

import

{

Intencion

}

from "@intencion/sdk";
import Anthropic from "@anthropic-ai/sdk";

const ix = new Intencion(

{

apiKey

}

);
const anthropic = ix.instrumentAnthropic(new Anthropic());

// every call captured: model, tokens, latency, outcome.

tools.ts

await ix.run(

{

intent: "refund_request", input

}

, async (run) =>

{
  const order = await run.tool("lookup_order", "orders-db", () => lookupOrder(id));
  return await issueRefund(order); // returns → success
}

);

Make your agent better, every week.

Start free in a minute. No call required, or try the live demo first.