Blog
/
Stigg Announcements

Reserve & Settle and Governance: Spend Control for the Agent Era

Reserve & Settle and Governance give you control over AI spend at every level, from each agent run to every team, user, and agent in your customer's org.

Or ArnonOr Arnon
Written by
Or Arnon
Last updated
November 18, 2026
read time
5
minutes
Many requests converge on a single budget check, and only some continue

Table of contents

Agents decide what to spend at runtime, many at once, and the cost is known only after the work is done. Today we're launching Reserve & Settle and making Governance generally available, so you can control spend before the work runs, not after the invoice lands.

Together they give you control over AI spend at every level. Reserve & Settle works at the action level: it holds the most a task can cost before it runs, settles the real cost after, and never overdraws, even with many agents running in parallel. Governance works at the org level: a budget for every team, user, and agent, enforced on every call.

Agents Spend at Runtime. Most Controls Find Out Afterward.

Compute is becoming the variable that decides how much work gets done. Research, testing, and data processing used to be limited by how much time people could put in. Now output scales with tokens, and agents decide how many to spend. Most controls weren't built for that, and two gaps show up as soon as agents reach production:

  • Spend is counted after it happens. An LLM call's cost is known only when generation ends, and an agent run's only after its last step. A balance check before the run holds nothing, so parallel requests pass the same check and together spend more than the customer has.
  • Spend is controlled in one place, but it happens in many. A B2B customer has orgs, teams, users, and agents drawing on the same credits. One agent stuck in a loop can drain the shared pool for everyone else, and nobody can say which agent spent it.

Either way, your customer can spend more than you charge them, and you eat the difference. Teams cover the gap with glue code: they overestimate and refund the excess, accept the overdraft, or build their own locking layer in front of the balance.

Control at the Action Level: Reserve Before the Work, Settle After

Every AI request is a spend decision. Now you can make it before the work starts, even when you don't know the cost yet:

  • Reserve. Before an LLM call or agent run, your app asks Stigg to lock its maximum cost. If the balance covers it, no other request can touch those credits. If not, Stigg denies it with InsufficientCredits.
  • Run. The task runs on its locked credits, whether it takes seconds or hours. They stay held until your app settles or the reservation expires.
  • Settle. When the task ends, your app sends Stigg the amount it actually consumed. Stigg deducts that amount and returns the rest of the hold to the balance, so the customer pays only for what was used. If the task fails, sending zero returns the full hold.

The reservation rides on the entitlement check you already make, with a reserveAmount added. Settling is a single call to the release endpoint with the consumedAmount, deducted to the same grants that backed the hold. You don’t build any reservation, refund, or reconciliation logic yourself.

Reserve and Settle flow: reserving 100 credits moves them from available to reserved, the task uses 40 while it runs, and settling returns the other 60 to the available balance

Concurrency: Many Agents, One Balance, No Overdraft

A balance check on its own can't stop parallel requests. If three agents each check a 100-credit balance for 60 credits at the same moment, all three pass, and together they spend 180.

With a reservation, the check and the hold happen together. Each granted reservation takes its credits off the available balance immediately, so the next request sees only what's actually left. A lock under the hood guarantees reservations never add up to more than the balance. Of those three agents, the first gets through and the other two are denied, since only 40 credits are left. The balance never goes below zero. When nothing is left, the rest are denied.

Three agents reserve 60 credits each from a shared 100-credit balance at the same moment: one is granted and the other two are denied

Safety: Nothing Stays Locked, Nothing Counts Twice

Reservations live on the same ledger as grants and consumption, so credits stay atomic, auditable, and safe under concurrency, even when a task crashes, retries, or runs long:

  • Reservations expire. A hold lasts one hour by default and up to 24 hours. When it expires, all its credits return to the balance and no usage is recorded, so an abandoned job never leaves them locked.
  • Retries are safe. The reservationId makes both steps idempotent. A retried reserve holds the credits once, and a retried release applies once.
  • Every step is in the ledger. Reservations, releases, and expirations are recorded in the customer's credit ledger and linked by reservation ID. Reserved credits show next to used ones on the customer's Credits tab, so finance and support teams can reconcile balances against ledger activity.

Control at the Org Level: Your Org Chart Is Your Budget

Governance, now generally available, lets your enterprise customers control AI spend at any granularity. Departments, teams, individual users, and AI agents each get their own budgets and limits, checked before every request runs, so no single user or agent can exhaust the shared pool.

  • Budgets that mirror the org. Customers define their own structure: departments, teams, users, agents, cost centers, or workspaces. Budgets cascade down, usage rolls up, and a parent's limit is the ceiling for everything under it.
  • Any dimension, any combination. Set budgets by entity, and also by model, feature, product, or region.
  • Know why a request was denied. The decision names the limit that was reached and the remedy: raise a budget or top up.
  • Cost attribution. Live spend by team, user, model, and feature that you can query, alert on, and share with finance.
  • Admins manage their own budgets. Your customers' admins set budgets and alerts themselves, without support tickets.
A Governance check passes the agent and user levels and is denied at the team level for insufficient budget

Get Started

Until now, agents decided what to spend and you found out afterward. Stigg hands control back to you, at every level: Reserve & Settle holds each action's cost before it runs, and Governance keeps every team, user, and agent within a budget your customers set.

Ready to take control of AI spend? Connect Stigg in one command with npx @stigg/terminal, or talk to us for a tailored demo.

Prefer to read first? See the Reserve & Settle docs and Governance docs.

Latest news.

One email per month.
From engineers, for engineers.

Thank you! Your submission has been received.
Oops! Something went wrong while submitting the form.