Going Rate

An agent should know the price of a job before it asks you to pay. Before you tap your passkey on a Kite spending session, Going Rate shows what people usually approve for that kind of task, so the agent asks for the right caps the first time and the run doesn't stop halfway for a second approval.

SERV ReasoningKite Passport sessionsThis page: passkey simulated

Agentdrafts an ask: task summary + caps

SERV ReasoningOpenServ inference API · classify + explain · gpt-6-luna

Going Ratecode: p25 / median / p75, cue, suggested caps

Kite Passportgoing rate in --task-summary · passkey · pay

Ask 1 · a wild cap

An agent asks for far more than the task needs

Its caps are a guess. Without a going rate, you either approve too much or reject and wait for it to ask again.

Per payment ($) Total ($) Duration (h)

Ask 2 · a similar task

This agent checks the going rate before asking

SERV puts the task in the same category as ask 1, then the agent asks for caps near what people usually approve for it.

Per payment ($) Total ($) Duration (h)

Ask 3 · too small

An under-ask that would stall mid-run

A cap below what the task usually needs gets a LOW cue. The agent would hit session_total_exceeded (or session_rule_exceeded for one big payment) and have to ask again. See it happen on Kite.

Per payment ($) Total ($) Duration (h)

Run an ask on the left.
The approval card appears here, before the passkey.

The rate card

Median caps people approved, by task type, plus what real Kite payments of that type were charged per call (ledger). Approval counts are split into seeded demo rows, your clicks on this page, and approvals confirmed through Kite.

Loading…

How it works

1. Agent drafts an askA task summary plus the caps for kpass session create: per payment, total, duration.

2. SERV classifies itThe OpenServ inference API, with a fixed system prompt compiled once into a cached flowchart, puts the summary into one of 8 task types.

3. Code computes the rateMedians and percentiles of past approvals for that type, compared with the ask: in line, high or low, plus suggested caps. SERV then explains the card in one sentence, using only numbers code gave it.

4. You approve onceThe going rate goes into Kite's --task-summary, so it's on Kite's own approval screen. After your decision, kpass session status records the outcome.

What's real here: two live SERV Reasoning calls per card (gpt-6-luna: classify, then explain), the rate math, the Kite-verified session decisions, and every payment on the ledger.
What's simulated on this page: the passkey. The going-rate CLI makes the real kpass session create call, waits on kpass session status, pays with kpass session execute, and records Kite's actual result. The contrast demo shows one such run.
What's seeded: the starting pool (93 rows, labelled seeded: true). kpass activity only covers one account, so a shared pool of real approvals doesn't exist yet. Going Rate starts one.