Agent Stream

Stop teams eating the budget

Stop one team’s runaway usage from becoming an end-of-month bill nobody can explain.

When every caller shares one provider key, model spend is a single number that nobody can break down. One runaway integration or a single busy team can consume the budget, and you often only find out when the invoice arrives. Without per-caller attribution, there is no way to cap usage in advance or attribute it to the department that incurred it.

What you can do

  • Set monthly USD budgets and token caps per member, team, or surface so that a runaway integration is stopped before it produces a large bill.
  • Set per-minute request and token limits per member, and a requests-per-minute ceiling on a surface, so that one caller’s burst of traffic is throttled before it exhausts shared capacity.
  • Attribute every token and dollar to a member and their teams so that you can answer who spent what, on which model, and charge AI spend back to the team that incurred it.
  • Route cheaper requests to smaller models and reserve frontier models for the hard ones so that you right-size the cost of every call.
  • Get alerted on cost spikes and threshold breaches so that a sudden anomaly reaches the people who can act, not just a dashboard nobody is watching.

How it works

MEMBERS & TEAMSTeam AMemberMemberTeam BMemberModelprovidersAgent Stream surfaceTEAM GATEResolve memberModel allow-listPer-minute limitsMonthly ceilingchecked in order✕ 403 · model not allowed✕ 429 · limit or cap reachedPER-STAGE METERINGLLM calltokens · $Judgetokens · $Jurytokens · $Decidertokens · $attributed to member and teamBreach & cost-spike alertsto your integrations

On a surface with team attribution on, each caller resolves to a Member, or to the built-in anonymous Member when it can’t be matched. A Member acts as a virtual key carrying a model allow-list, per-minute request and token limits, and monthly ceilings. Members are grouped into Teams, and once a surface is set to contribute to team quotas, spend and tokens are attributed to both the Member and its Team across every eligible surface. Each Team’s Usage tab shows what it has spent.

On each request the Team Gate resolves the caller, checks the model allow-list, then enforces per-minute limits and monthly ceilings before the provider is called. Teams carry monthly ceilings only, while per-minute limits sit on each Member and, through the Rate Limit element, on the surface as a whole. Cost, tokens, and latency are metered per stage, so the LLM call, the Judge, the Jury, and the Decider each show their own share rather than one combined figure. With Mode set to Decide LLM based on content, the Decider reads each request and sends simple ones to a cheaper model and complex ones to a frontier model.

Usage limits and breach alerts are separate mechanisms: a limit is a hard ceiling, while a breach alert only notifies. Once spend or tokens reach a monthly ceiling, further requests are rejected with a 429 before they reach the provider, and the request that crosses the ceiling is itself allowed through. A request over a member’s per-minute limit is also rejected with a 429, and an access or model-allow-list failure returns a 403.