FinOps for AI Agents: Who Spent All the Tokens? — Tisha Chawla & Susheem Koul, Microsoft

AI Engineer · 21 min · 212 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:13Okay.
00:14Um, good morning everyone.
00:15So, um, I'm Tisha and I have Sushim with me as my co-presenter.
00:20All right.
00:20So, we'll be talking about the most expensive question in AI today.
00:26I think a lot of you would have come across the scenario that um you know when you opened an AI bill like through your agent workflows um you couldn't actually trace back where that bill was actually coming from right
00:40and um and I don't think that's a problem right now because right now the industry is valuing you know um token maxing that is like spending the most amount of tokens for exploration
00:53for all of those purposes And um people are proud to call themselves token billionaires and um I think that's all right but this talk is you know the shift from token maxing
01:05to value maxing you know how do we get there and um we'll talk about it from this question um who spent all the tokens and um if anyone spent all the tokens there has to be value associated with this right and that
01:24is um the talk about.
01:30All right.
01:30Now in order to minimize the gap you know from token maxing to value maxing we'll kind of see we'll observe the patterns which the like the existing um u like the past software evolution eras had
01:47like for instance when we talk about the SAS era the interface was UI and the control was in the form of usage caps right like or the seat limits or tier based policies
02:00Now when we moved on to the cloud era, the control surface again changed.
02:06The model became pay as you go and the control moved like in the form of autoprovisioning and you know autoscaling policies.
02:14Now we are in the agentic era right and um now how the cost is calculated here is in the form of model calls right like u how like the
02:29code calls your model but what we've observed is that there isn't a proper control plane in place for that like we do have control plane in place for you in place as model gateways where
02:43they're um are hard caps or there is model routing to downgrade the model but the part like where the code you know calls the model that um is what we'll be talking about
02:58uh today and um we also you know see um like in the last year we've seen a lot of unbounded consumption happening like um if you've read the news.
03:14There was news about the like the uh AI budget for Uber getting exhausted within 4 months and um there were companies who like who ran into you know like hundreds of millions of dollars within just months or days and like there were a lot of
03:31um like other news in place as well where like these runaway loops um led to a very like massive increase in the cost and there wasn't proper mechanisms to control it.
03:47Um so when we see all of this the first thing that comes to our mind is is there a tool or is there a product to save us?
03:54But uh we'll instead talk about the first principles of how you know we can design a system which is actually true enough to solve the problem from the very root.
04:07So for that let's um like dive onto the principles.
04:12First of all, let's talk about token being the unit of cost.
04:16Right?
04:16We are charged in terms of token.
04:19So the now we have to see value also in terms of token.
04:24Right?
04:25Next um we all know that cost is created at the LLM like the model call boundary.
04:32Um so that is what we'll have to track and if we don't have proper attribution like if we don't know what agent want run made that particular call we we can't you know control it right we we just know the like the broad
04:50uh picture of what went wrong but we don't we can't you know trace it back or narrow it down.
04:55So that is why attribution is a very important element to have and um like once you know which particular run or which particular agent is actually you know attributing to the cost you should have proper policies in place to actually stop it.
05:13Like um let's take example that um if you have a you know um a loop which is you know running um very excessively and which is not required or you know if your context is growing very out of range.
05:29you should have in place policies which can um like solve that particular thing there and there instead of halting that and if um and as the last resort only a like a halting or a
05:42should happen from a budget cap.
05:45So these are the first principles.
05:46Now let's see how we can you know define an ideal uh platform on top of that from these principles which we talked about.
05:56All right.
05:56Uh so one thing which is very important that which matters here is that u when we talk about um like the um existing frameworks for token ops or for token management
06:08most of them are at the um like u basically monitor the model uh request.
06:16they like they are like model gateways which will u you know u basically um do like model routing or hard budget capping.
06:27But what we need right now is something which you know um like monitors you at the run instead.
06:35Like um if you see we need something uh which can control the loop between like the agent call between the tool um and the agent.
06:48something you know which can um uh see or control the the spawning of multiple sub aents happening from a one main agent or um like something which can control the growing of context.
07:02So like that is the need of the right and that is what we need.
07:06So for all of this um we like uh kind of are proposing a platform which first of all um has a cumulative budget across like the uh like the
07:20attribution runs which happened and then where enforcement actually happens in call path rather than um you know a separate thing like for example if something goes wrong if your like if your context is just growing heavily.
07:38Then like in place compaction should happen or like in place caching or something like that should happen.
07:47And um after that if like after basically exhausting the list of all in place policies only like uh the budget cap should happen at the very last.
07:59Um so that is something which we are proposing.
08:02But um if you look at the landscape today, if you see the uh like the uh tools like um this light LLM, port key, cloudflare, all of those they happen
08:13at again the request level right um like if you see like halting is there, routing is there for some of those but all of this again is at a request and you can't control the cost at the uh request layer
08:30uh at the model layer, Right.
08:34So this is the missing piece which is you know the u basically navigating it at the um you know the model the agent run layer.
08:47So for that we have token ops which is uh you know a runaway token governance for AI agents and u this is the uh architecture for that.
08:58So first of all one thing I would want to highlight is the like the intentional design decision we took here was an out ofbound plane.
09:07So it doesn't interfere with your code at all.
09:11Um so if you see here that out of the bandound plane has three modules which I'll be talking about.
09:16The first one being instrumentation.
09:18It is a common observability layer where you know you'll u like uh have u like the basic telemetry the open telemetry the cost in microns and um like the
09:31um like enrichment layer and basically um the uh attribution like what caused that uh like particular run and then there is um obviously accounting where you'll basically accumulate it in a kind of a ledger
09:52like the total runs which are happening.
09:55And finally we have this enforced layer which has uh two main purposes.
10:01one is steering it um through the policies which we've defined which I think will cover later and um then we have halt in place as the you know final
10:13um like u final thing if um you know your budget is getting exhausted so yeah that is there now when we again look at the landscape this kind of will
10:26solve a lot of problems um which kind of happened uh when we like look at the previous um tools or products there because uh it is at happening at run and it is you know uh helping you solve the problem from the very root by steering
10:46it in place.
10:48All right.
10:49So uh with this I would like to hand it over to Sashim for the demo.
10:53Yeah.
10:57Oh yeah.
10:57Now I think I should be able to everyone in the back can hear me.
11:00All right, perfect.
11:01So yeah, we have established the principles behind token ops till now.
11:06Right.
11:06Now let's shift gears, talk about the design part of it and uh maybe get into the code and the eventual demo.
11:13Right?
11:13So what I have behind me on the screen is the like bird's eye view of what token ops looks like today.
11:20It's it's three layers.
11:21We'll go left to right and top to bottom.
11:23So on the left most you have your own agent runtime which you're trying to instrument and kind of manage the cost for right the middle layer is what we're calling the bridge that
11:32basically shuffles data between your agent and the control plane and the control plane is where the mind of the system lies right so let's talk about the bridge layer very briefly
11:41if we uh go from top to bottom you have the attribution on top so what we're trying to do here is every agent run that you do it's attributed to some user dimensions so the idea is everything that you do every run of the agent is accounted to some usability or some usage.
11:57This comes in handy later.
11:58We'll talk about it.
11:59Uh the second part which is the boundary annotation that you see this is pretty much the heart and soul of this middle layer.
12:06So the idea behind the boundary annotation is that you take any method.
12:09It doesn't matter what framework you're using.
12:11You might be using uh let's say lang chain lang whatever.
12:14If you have a method you can annotate it with boundary.
12:17What this annotation is going to do is it's going to do two things.
12:19First it's going to track the input and the output and it's going to flight that up to the control layer and record it there as a ledger entry.
12:27Now this will be annotated with the further agent run ID and the other attributes and so on.
12:31The second thing the boundary annotation does is it acts as a channel through which the control plane can push actions down to the agent.
12:38This is where the entire intelligence lies.
12:40So we do not have a single directional highway.
12:42We want the control plane to be able to tweak the behavior of the agent on the fly to ensure that we are able to squeeze in more runs inside our budget cap.
12:51Right now let's say the control plane pushes down an action.
12:54Let's take a small example.
12:56Let's say you have a rag retrieval tool which is generating like 20 chunks every retrieval for every call and that's eating up eating up your budget.
13:03And let's say the LLM is not even using the chunks that are after five because they are just not relevant, right?
13:07They're sorted by relevance.
13:09So let's say the control plane observes this and it wants to limit the output to just five chunks.
13:14So it can push down an action but that action has to be received by boundary and then has to be executed by something.
13:20That is where the third node, the governor node comes in.
13:22The governor knows what actions are allowed on your agent by you as a developer and it receives those actions from the control plane and knows how to apply it in a non-destructive way.
13:33So that's the first three.
13:35The fourth one wrap uh the wrap complete is essentially just a helper method.
13:39So as we know most of the agent providers or the model providers they provide objects rather than methods for their LMS right.
13:45So wrap complete is just another way of applying boundary on objects rather than methods.
13:49Let's shift right to the control plane.
13:52On the control plane the first layer is the segment.
13:54Now this is where the attribution that we talked about earlier comes into picture.
13:58So any dimensions that you float from the attribution layer.
14:01Let's say you have a preview agent that you share with everyone in this room and your agent is floating a dimension saying that cohort is AIE 2026 right so you can create a segment which is a cohort of users
14:14which is based on this tag like dimension being AI 2026 right and you can apply your budgets at this cohort level so you don't necessarily have to restrict everything at an agent level or a run level you can do you can do rollups you can do fine grain or coarse grain control right so that's the segmentation part
14:31of Ledger as I mentioned is just one agent run all the traces in one place.
14:36Then you have budgets.
14:37Budgets are basically just the static thresholds that work across a time window against a particular segment or an agent run.
14:43And then you have actions.
14:45So on the actions part we have broadly two flavors.
14:47First is the halt type actions which basically just kill your agent if it exceeds a budget.
14:52The second part where we are adding value is the steer type actions.
14:56So here we do not kill the agent.
14:58Instead we try to steer the behavior of the agent or the components of the agent to try and fit that particular run within the alerted budget.
15:07Right?
15:07And then the policies layer is where it all comes together.
15:10You basically uh group the budgets the actions and then set your policies against certain segments or agent runs and that is where it executes.
15:17Right?
15:18So moving on uh what changes in your code that is the boundary annotation that we just talked about.
15:23As Disha mentioned earlier this is all out of band.
15:25So you do not have to change your code.
15:27You just have to apply the annotation on the methods that you have.
15:29This boundary annotation will take care of floating all the information up to the control plane.
15:33And uh the control plane lies in your own tenant.
15:36So you do not need to worry about any data leaks or anything.
15:39Then if I talk about the governor, so for the governor, you just have to create an instance.
15:43You just have to pass it your own configs.
15:45These configs will basically declare what sort of actions are allowed for those agents, right?
15:50so that your control plane cannot just willingly do any random things on your on your agents.
15:56So before we move on to the demo, I'll just briefly touch upon the uh test that we're going to use today.
16:02So it's a simple two agent workflow.
16:04We have a research agent which has access to a search tool.
16:07Uh you give it a question.
16:08It's allowed to look up on the web as many times as it wants.
16:12And once it knows that it has all the data, it passes the findings on to the second agent which is a summarizer which creates creates a research report.
16:19Right?
16:20So with that out of the way, let's just quickly walk over to the demo.
16:24So for the demo, we have three different scenarios that we're going to talk about.
16:27For the first one, we're going to run the token ops in what we call preview mode.
16:31So in preview mode, what happens is that all the policies run as is, but the enforcement doesn't happen.
16:36So if you see we ran a particular run over here which completed but we did not see any sort of failures there.
16:46The policies executed but the actions that were associated with those policies were not allowed to be executed.
16:52So we're just going to load the dashboard screen here.
16:57Yeah.
16:57So this is the governance output.
17:00Governance is off.
17:00The run completed.
17:02But in the dashboard you can see the policies have executed.
17:04So you can see the cost budget, the cost guard and so on.
17:07Right?
17:08So this was the first scenario.
17:09For the second scenario, what we're going to do is we're going to turn on the governance.
17:12Now while that is happening, I just want to touch upon why this is important.
17:16So if you want to like include this product into your production agents, you want to have a safe environment or a safe way to firstly put it in your production environment,
17:24test the guardrails, tweak the guardrail, see what's the policies are doing and then finalize the thresholds.
17:31Right?
17:31So this is the second one where we have now enforced the governance and you can see in the dashboard that the pre-all cost cap has exceeded.
17:39So you had a budget allotted for this run but the agent exceeded the budget and it was killed immediately.
17:45So that's the simple circuit breaker sort of a methodology.
17:48So this is the halt behavior.
17:50And now let's see the steer behavior which is the which is where we are trying to add value to this entire cost management scenario.
17:58So this time we're going to run the third the second prompt.
18:01The budget allotted for this one is slightly higher but it's still not high enough for the agent to complete in time.
18:07So what instead happens is there is something called cost guard which kicks in.
18:12This cost guard it takes into account two things.
18:14First how much of your allotted budget have you consumed?
18:17Second what is the velocity at which you're consuming tokens.
18:20[music]
18:20Now based on these two things if it predicts that you're going to run out of your tokens or your allotted budget by the end of the run it's going to inject something into your system instructions
18:30that something could be as simple as hey you're running out of budget so make sure that the LM outputs are more succinct or more summarized right so that is the
18:40way we are doing the steering now the this was a very simple test bench to show you like how this works on a like working code we have also benchmarked it on a couple of open source repos.
18:52So we have benchmarked it on browser use as well as metagp.
18:55Uh we ran it across multiple iterations across stress tests across simple scenarios hard scenarios and everything.
19:01And the results we see are the average spend goes down by almost 78% with token ops enabled with the full policy suit that we have today.
19:08On the completion part when we compare it with throttling just simple throttling your simple throttling is going to kill your agent runs no matter what.
19:16Right?
19:17So with the reduced average spend what you get is you get an uplift in that completion percentage from 67% to roughly 96%.
19:24So that is the value ad that token ops is doing here.
19:27Now this is the policy catalog that we run this benchmark against.
19:30This is what we support today.
19:32We kind of researched what are the different failure modes that are there today out in the wild and tried to cover most of them here.
19:38So you have things across spend management, you have things across context management like context compaction, tool output reduction, you have things across loop detection and progress detection and stuff like that.
19:47So this is the entire set of policies that we support.
19:49And at the bottom you can see the actions.
19:51So as I mentioned earlier, we have two flavors.
19:53You have the uh the halt type actions and then the steer type actions.
19:58So for the steer we can do allow, mutate, inject and so on.
20:01And for the halt, it can be a simple kill.
20:03But this is not the end state that we envision for this.
20:07The end state is we have a lot of data right we have a ledger that is continuously being updated.
20:12So what we want to try is we want to try a self-learning module within the token ops plane within the control plane which can look at this ledger and ask this question hey why or what is the failure mode that I'm still not able to catch and then based on that it can do two things one is
20:29it can enhance it can generate new policies on the fly based on the missing or the still uh runaway costs or it can refine the existing parameters for the existing policies that are there so that the runaway costs are managed more effectively in the future.
20:45So with that I think uh that is all we have for you guys today.
20:49Thank you so much for your time and you can scan this QR code that's the public wiki.
20:53We are updating it almost regularly.
20:55So you can scan this and stay up to date and uh Tisha and I are around so if you guys have any questions or if you want to discuss more about it just let us know.
21:03That's it.
21:03Thank you.
21:05[applause]