A cap on the spreadsheet fan-out.
=AOP dragged down 500 rows with a $25 budget. At $25 the run pauses and asks. Usually it finishes under.
Trust · Athena Governance System · Hard Budgets Per Run
Control agent spend with hard budgets per run
Cap every workflow in dollars or model calls. The runtime stops at the limit, tracks cost by workflow and user, and fails closed if the ledger is unreachable.
Two smoke runs cost $200 and $101. Then we built budgets.
- Hard budgets per run
- Atomic across sub-agents
- Fail-closed ledger
- Cost by workflow, agent, user, project
- Stale and duplicate detection
- Every middleware observable
The problem
Autonomy without limits is a liability
Agents that fan out do real work and spend real money. A spreadsheet formula dragged down 500 rows is 500 runs. A sub-agent per audit sample is 125 concurrent runs. Without a hard cap, "the agent got stuck in a loop" is a finance conversation.
A monthly invoice rarely shows per-run detail.
Autonomy with a limit is a tool. Without one, it is a liability.
How it works
Set. Enforce. Watch. Prune.
Set. Per run:
max_cost_usd,max_model_calls, or both. Per user, per project, per Skill (reusable component) on the dashboard. Set it in the AOP editor or in one field on the API call.Enforce. The runtime reserves cost before each model call in a shared transaction ledger—a real-time spend tracker that concurrent sub-agents read and update atomically so fan-out cannot overspend. If the ledger is unreachable, the run fails closed.
Watch. Cost by workflow, agent, user, and project. Run steps, exports, and a per-run summary written by the run itself.
Prune. Stale and duplicate procedures flagged. Role-based visibility into who can see and run which procedures. Approval and cost estimation required for high-volume sub-agent fan-out.
In practice
Budgets and observability in production
Cost per user, per practice group.
A team asks whether an internal knowledge bot is worth it. The dashboard answers by user and by group.
Every governance layer signals when it runs, enforced by the build.
Each governance layer emits one production signal every time it runs. A test fails the build if any middleware goes silent, so a "silent no-op" regression cannot ship.
Stale and duplicate detection.
Procedures nobody has run in 60 days, or two that do the same thing, surface for review. Retire or merge.
Approval before the fan-out.
High-volume sub-agent runs show a cost estimate and require approval before they start.
What customers say
“It's good to understand the capability of the platform. That how far, and how quickly you guys can get these things going [...] that has been impressive overall.”
“Lets us get a lot more out of our data than [...] We can sit on top of all of the data in [our data warehouse] as opposed to just, like, our little ingested, you know, three gigabytes semantic models. I think we have a lot of power here.”
“Gave Athena the spreadsheet, explained which column I was trying to figure out, and then ten secs it told me... exactly how the calculation was based on the other data.”
“He wrote the entire stuff. By itself. And I was there, and I was thinking, alright. If I had to write all this stuff, I mean, I would've spent days, days, days.”
“I send that to the programmer and he said, alright. I read your file. I copied the section of the code because he has the right interface, and it works. Wow. Just like magic.”
“One workspace. All different, let's say, avatars of their output data. Right? Suddenly, it's dashboard. Excel is here. PowerPoint is when it's all together. It's very impressive.”
“Use the Athena platform probably closer to the concept and how it was designed, like, with the spaces architecture where you've got a set of primitives and capabilities, and you bring together the right set of capabilities to solve a task.”
“Digital worker, it gets spun up, it gets given access to the toolkits that it needs. It gets given access to the data that it needs. Just exists for the period it needs to exist to perform the task. And then it gets torn back down again afterwards.”
“I like that direction, that concept of the agent is the agent. It becomes an entity in its own right with its own ownership and its own missioning, and you get rid of the gray area and the blurry lines.”
“It was PowerPoint... there was no kind of distinction between the two. So that was killer. And I literally had a meeting on it today, and people said how great it was.”
Verbatim from customer calls. Customers anonymized.
Stated plainly
What budgets and governance do
Hard budgets per run; shared ledger across sub-agents; fail-closed.
Dashboards by workflow, agent, user, project; exports.
Approval and estimation on high-volume fan-out.
Observability enforced in CI.
What to plan for
A budget stops spend. It does not finish the job. Set it with headroom, and read the run summary when it pauses.
FAQ
How do we control AI spending?
Set hard budgets per run, workflow, and workspace; runs stop when a budget is reached.
Can we see where usage goes?
Yes, by user, team, and workflow.
Who sets the rules?
Your admins control tools, models, connectors, and approvals.
The rest of the platform
Related products and stories
Pick the workflow you were afraid to let loose.
We will set a cap and let it run.