WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=TeErpYBUIeM

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/TeErpYBUIeM.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.559 --> 00:00:14.799
Everyone, super excited to be here.

s3
00:00:14.799 --> 00:00:16.560
Ah, okay.

s4
00:00:16.560 --> 00:00:18.640
My laptop was alerting me that this starts now.

s5
00:00:18.640 --> 00:00:19.439
Uh, hi everyone.

s6
00:00:19.439 --> 00:00:21.840
I'm Mark, one of the founders of Langfuse.

s7
00:00:21.840 --> 00:00:29.920
super excited to chat about uh what we have seen how how people like self-improve agents now and what kind of like open source reference stack uh we see emerging.

s8
00:00:29.920 --> 00:00:40.559
I mean obviously my view is based on working with our community of people that use language to trace and evaluate their their agents but I think I try to keep the talk mostly to to like more generic takeaways

s9
00:00:40.559 --> 00:00:42.879
that work with whatever stacks you're using.

s10
00:00:42.879 --> 00:00:46.960
Um but yeah happy to chat about the different pros and cons after the talk.

s11
00:00:46.960 --> 00:00:47.760
high level.

s12
00:00:47.760 --> 00:00:59.039
I think this is the year uh where uh teams talk about like upleveling the the level of uh where where they operate themselves where increasingly good good models just abstract everything downstream.

s13
00:00:59.039 --> 00:01:05.360
I mean coming from I don't write prompts I have loops u Peter writing about it auto research becoming really popular.

s14
00:01:05.360 --> 00:01:15.280
So I think this is kind of where things are going on the like like producing um application side but also on the how to actually um build good good agent applications

s15
00:01:15.280 --> 00:01:33.119
and um over the years really models have have expanded a lot in capabilities um back when when we started working on length fuse this was right when ch was released so uh where like even the simplest applications didn't really work so for example we built like github issue to pull request automation similar to what devon or kurs

s16
00:01:33.119 --> 00:01:39.759
are building but back then we could only manipulate like single page applications like a single HTML file because multifile edits were too complicated.

s17
00:01:39.759 --> 00:01:44.560
Since then all of these things really accelerated a lot and this really made made loops not possible.

s18
00:01:44.560 --> 00:01:53.200
At the same time a more like reference tech emerged of how great AI agents are built because building agents is different from building applications.

s19
00:01:53.200 --> 00:02:05.600
It's like an like a link of like tracing and monitoring like online how users really use an agent application and then like an offline component of how to build data sets how to experiment

s20
00:02:05.600 --> 00:02:16.640
uh regarding like making changes to these agents and then evaluating offline whether things work then deploying new to production then learning again for real users use the application uh so it's like this this kind of loop where traditionally

s21
00:02:16.640 --> 00:02:30.720
I mean you have like obsibility and the data docks here in the tracing monitoring side and you have like the ML ops toolings like MF flow weights and biases more on the like offline side where this new category emerged and that's also why we built length fuse because you need to bring the online and the offline together

s22
00:02:30.720 --> 00:02:44.080
as as otherwise like you either benchmark on data that's inaccurate so the data sets they're not in loop and like uh not in sync with what's happening in production or you monitoring like production data but you don't really benchmark offline so you need to bring the the both together

s23
00:02:44.080 --> 00:02:53.840
but this has been like a lot of manual labor because you need to look at traces update these data sets think about like new evaluators uh then make these changes create new hypothesis

s24
00:02:53.840 --> 00:03:08.080
like it's it's like a lot of work going into this process and since starting working on language we try to educate teams how to how to do this process well but it's it's usually like the ask them like oh models got better now how can we take ourselves out of the loop to uh to automate it because

s25
00:03:08.080 --> 00:03:17.440
it's really tedious uh to get things going and that's what I want to talk about how we see how people go from manually driving this loop to using AI to to drive this loop

s26
00:03:17.440 --> 00:03:26.640
um we use the visualization um like of for example like the loopcraft article or like many many articles in the space that like layer different loops on top of each other.

s27
00:03:26.640 --> 00:03:33.760
How I mean at the at the lowest level you just have like this token loop of what's the next what's the next token that's produced and on the very high level it's

s28
00:03:33.760 --> 00:03:38.319
uh like like the most abstract form of thinking where you as a human are still involved.

s29
00:03:38.319 --> 00:03:46.560
So what we see how um we go through them one by one high level I just want to highlight these are again possible because models improve.

s30
00:03:46.560 --> 00:03:57.280
So uh when when we started working on length in 2023 like I mean we had github copilot so like the lowest level or the uh level two u then level three was possible like more in 2024

s31
00:03:57.280 --> 00:04:07.519
and now2526 is more like the higher level loops of actually agents reasoning about what what to even fix uh how to propose new fixes and how to how to verify whether they actually work.

s32
00:04:07.519 --> 00:04:15.920
Where humans are currently still involved is one you need to align data sets of example questions of how the agents are used.

s33
00:04:15.920 --> 00:04:24.160
This is so like where where humans need to review um like like what what gets amended to the data set and how to then evaluate whether the agents actually work.

s34
00:04:24.160 --> 00:04:26.880
These are the the two blue things in in loop number three.

s35
00:04:26.880 --> 00:04:37.440
Then on the proposal of fixes because if um like your agent proposes how to change an agent, you still want to see what was changed because you have the risk of like overfitting it to

s36
00:04:37.440 --> 00:04:40.639
um to like the data set and the failure definition.

s37
00:04:40.639 --> 00:04:50.800
Um, so I mean usually agents find all sorts of different uh failure patterns, but maybe some of them don't really matter that much and you're like, huh, this kind of use case for the agent, it doesn't really matter that much.

s38
00:04:50.800 --> 00:04:53.280
Like this is like out of what the agent should be actually be doing.

s39
00:04:53.280 --> 00:04:59.840
So we don't need to over fit our application to this random quirk that we found in production, but we want to like be in the loop to to monitor this.

s40
00:04:59.840 --> 00:05:10.400
And now coming from this loop pattern, we try to dissect it again into more like a workflow where where we currently see AI used the most is in proposing fixes.

s41
00:05:10.400 --> 00:05:20.080
So if we have production traces and we have like evil criteria and data sets of how to reproduce issues that we find in production, then this how do we hill climb against the data sets?

s42
00:05:20.080 --> 00:05:27.039
That's like a really cool uh way to kind of like loop your way to a success with with agents because usually they're like 10 different things of what you could be doing.

s43
00:05:27.039 --> 00:05:31.840
I don't know, use a different model like aggregate context in a new different way.

s44
00:05:31.840 --> 00:05:38.560
Try whatever you find on X uh in a given week to to see whether it can like uh make a dent into improving the agent.

s45
00:05:38.560 --> 00:05:42.240
But really like you maintain the boundary of what you optimize against.

s46
00:05:42.240 --> 00:05:46.880
And you use AI mostly for like proposing improvements to the agent implementation.

s47
00:05:46.880 --> 00:05:55.840
Like over the last I'd say month, we see more and more teams also using uh like like AI and and their loops to maintain evil criteria and maintain data sets.

s48
00:05:55.840 --> 00:06:04.000
So uh usually what goes in here is you want to like align how a data set uh looks like for an agent with what actual users are doing.

s49
00:06:04.000 --> 00:06:14.560
So for example if we have an customer support application like support agent then the data set should be like common support questions but like users do all sorts of things with a support application.

s50
00:06:14.560 --> 00:06:20.080
So you need to continuously keep it in line with uh with what users are doing and you can use agents for that.

s51
00:06:20.080 --> 00:06:28.160
And for EVAs as well if you see usual error patterns then you can reproduce the error patterns and propose new evil criteria on these offline data sets.

s52
00:06:28.160 --> 00:06:30.960
So for example I don't know I want I don't want to name competitors.

s53
00:06:30.960 --> 00:06:31.919
I want to be concise.

s54
00:06:31.919 --> 00:06:34.960
I want to answer in the same language as the user actually requested.

s55
00:06:34.960 --> 00:06:37.360
Um like the question in my customer support application.

s56
00:06:37.360 --> 00:06:39.919
So you can also propose new evaluators.

s57
00:06:39.919 --> 00:06:43.280
This is I think rather newish over the last months.

s58
00:06:43.280 --> 00:06:46.319
um and where we see like the AI more autolooping.

s59
00:06:46.319 --> 00:06:55.199
However, usually teams that are still involved in reviewing these changes because they create the new boundary for how then other agents try to auto improve against it.

s60
00:06:55.199 --> 00:07:11.199
If you give this completely out of hand, you risk like that you just create more slop where we've used this in like a blog post of ours where I'd say the target of an AI application is always changing because you don't really know like usually you start with a very high level task of I stick with customer

s61
00:07:11.199 --> 00:07:11.680
support.

s62
00:07:11.680 --> 00:07:19.840
You stick with someone says hey in our company customers need to wait a long time for support answers and we have a lot of people employed to respond to them.

s63
00:07:19.840 --> 00:07:27.280
AI should be able to do this automatically or at least eight in customer support, but it's a very wake target because people don't even know what happens in customer support every day.

s64
00:07:27.280 --> 00:07:29.599
So, it's kind of like, oh, we want to automate all of it.

s65
00:07:29.599 --> 00:07:34.560
However, then on the way of doing this, you figure things out based on like error cases of what you even need to do.

s66
00:07:34.560 --> 00:07:38.560
And that's like this kind of like map that you like you assemble the plane while you're flying it.

s67
00:07:38.560 --> 00:07:49.199
Um and how we then look at the uh at this kind of like looping graphic is you really want to be like involved in setting the setting like the the goals

s68
00:07:49.199 --> 00:08:05.120
and how you make changes to data sets evaluators you want to be in that loop involved because you set there by the direction and then AI can automate against it and then off of the newly implemented changes you'll see new error classes and you can then like change cause again on the most highest level but thereby you

s69
00:08:05.120 --> 00:08:11.440
pull yourself out of a lot of like the more manual and tedious work on the lower levels.

s70
00:08:11.440 --> 00:08:17.440
Um, this is then the high level view of how how I would conceptualize like the different scenarios that you can find yourself in.

s71
00:08:17.440 --> 00:08:29.919
I'd say if you implement this uh like AI workflow well of tracing online uh how your agents are working implementing evals on how they are working um having data sets

s72
00:08:29.919 --> 00:08:44.159
uh of example questions and then benchmarking against them you are in this era in this class like very high time invest because you need a team who really works with this data every week but also quality of the of the agents is pretty high I mean that's what teams have been for example using length use for for

s73
00:08:44.159 --> 00:08:48.720
the last years and this is how you can get to I'd say success with your agent application.

s74
00:08:48.720 --> 00:09:01.920
Um, if you fully automate it, I would say you don't really need to invest any time because you're just like, I don't know, codeex goal mode your way to success or like you add higher level loops on top of it, but you also risk that like it produces a lot of tokens, but don't really make sense.

s75
00:09:01.920 --> 00:09:11.120
So, I think you want to be in this kind of like promised land of you don't really need to invest a lot of time, but you're like in the loop at the most important steps of like setting

s76
00:09:11.120 --> 00:09:15.600
the direction of where you want to go, but you take yourself out of everything that's tedious.

s77
00:09:15.600 --> 00:09:18.800
Thereby you're way less time invested than running it manually.

s78
00:09:18.800 --> 00:09:32.959
But also the quality can be even higher because usually if you need to do everything manually you're bottlenecked by your own time your own like perseverance endurance uh or like patience to look through so much data and AI can just look so through like so much more error cases.

s79
00:09:32.959 --> 00:09:36.320
This agent will be higher performing and you need less time invest.

s80
00:09:36.320 --> 00:09:48.000
That's the I think that's what we were going for and how this then looks like is uh you involved at the top but also I mean you want ideally your application to feed in interesting signal.

s81
00:09:48.000 --> 00:09:58.000
So for example when we talk about the like customer support application use case usually I mean customers can swear at the support agent customers can be like oh no you don't understand

s82
00:09:58.000 --> 00:10:05.279
uh like this was not correct or like you can propose messages to an internal agent who then can accept them or change them when sending them out.

s83
00:10:05.279 --> 00:10:12.240
So there's like a lot of like implicit signal that can feed into your process without you as the agent developer needing to do this manually.

s84
00:10:12.240 --> 00:10:17.040
Uh thereby you have like an like an instream of of signal that your loop can act on.

s85
00:10:17.040 --> 00:10:23.600
Um so it's either you your team or some signal coming in um and the lower level loops can be automated.

s86
00:10:23.600 --> 00:10:25.839
That that's high level how we how we think about it.

s87
00:10:25.839 --> 00:10:31.920
And like uh like we put together like a super quick demo application um that can that can show this.

s88
00:10:31.920 --> 00:10:43.760
Um so for example like we as the language team we have a big problem now because our like engineering team uses AI a whole lot thus they ship a whole lot and somehow we need to update our customers

s89
00:10:43.760 --> 00:10:51.760
uh and user base uh what is what is new every week because uh like in recent weeks like a lot was released uh and usually need to update documentation

s90
00:10:51.760 --> 00:11:01.680
update a change log post about it on socials and historically like engineers did this themselves because they maybe released like a bigger feature like once every month so it's okay to spend a couple of hours.

s91
00:11:01.680 --> 00:11:13.200
But now if they take themselves out and really ship a lot of like product, they want to ideally uh like automate also that kind of like release process because otherwise we are bottlenecked by how fast we can communicate about changes.

s92
00:11:13.200 --> 00:11:21.040
So um we thought about like putting together like a change lock writer that just uploads like like updates a lot of like the public assets based on what we've shipped.

s93
00:11:21.040 --> 00:11:31.120
Um however then the question is is like how we communicate good because if we communicate badly about what we've released then I mean we we kind of like sabotage our releases

s94
00:11:31.120 --> 00:11:34.560
if we like misrepresent for example what has even been released.

s95
00:11:34.560 --> 00:11:53.120
So um setup is we have a merge PR, we have a change writer that uh suggests a change to documentation, but then there's like a just like a code review step as like a GitHub PR where someone can review and either like approve on GitHub and get it merged or like submit like change requests of like what

s96
00:11:53.120 --> 00:11:58.560
needs changes in this um in this in this change in documentation and and the change log.

s97
00:11:58.560 --> 00:12:02.800
Then the change writer would act on this again to update the draft until it's finally released.

s98
00:12:02.800 --> 00:12:16.880
And now both of these kinds of like either like approvals or uh requests for changes can get feed into like the AI obserability and eval to to serve as a basis for auto improvement of that change writer agent.

s99
00:12:16.880 --> 00:12:33.680
So what we've now done is um just use a coding agent pointed at the like length use agent skills but generally I mean this would also work with other systems but yeah length is like well positioned for this to to ask like okay what has happened in this application how can we improve it I'll I'll pause a

s100
00:12:33.680 --> 00:12:36.720
couple times because otherwise it it'll go it'll go by very fast.

s101
00:12:36.720 --> 00:12:50.480
So what we for example here identified that the writer uh so the chain lock rider is factually uh like very correct in how it makes changes to documentation but clarity is low.

s102
00:12:50.480 --> 00:13:09.040
the edit ratio of us humans making ch like submitting change requests is still high and that we leak internal jargon to our customers because maybe our code has like internal comments that we leave there for ourselves for the future but uh like our customers don't think about our application as like I don't know like an ingestion pipeline

s103
00:13:09.040 --> 00:13:20.800
or as like I don't know some kind of like batched evil Q like they don't know this is like an implementation detail we should talk about it as like an user domain language and that's like uh something we have identified here we're Now

s104
00:13:20.800 --> 00:13:33.680
the agent suggests changes to the data sets and evaluators to test for leaking jargon and um also like increasing clarity in customer domain language basically.

s105
00:13:33.680 --> 00:13:53.600
So um agent now closes this gap suggest changes to the data sets and evaluators and now um has an idea of how to change the agent implementation to basically fix this problem because in the end it's kind of like just like a a problem of how we provide context how we like provide the skill that then educates

s106
00:13:53.600 --> 00:14:02.240
this um content writer and now we can back test it basically on the updated data set to see whether the new implementation would do better than the old implementation

s107
00:14:02.240 --> 00:14:06.639
and I'll stop here again where we have basically the baseline.

s108
00:14:06.639 --> 00:14:12.000
So um basically we reproduce the issue of userf facing language.

s109
00:14:12.000 --> 00:14:15.360
So let's talk about features in a way of how users would talk about it.

s110
00:14:15.360 --> 00:14:18.480
So we reproduce this kind of problem with the v1 baseline.

s111
00:14:18.480 --> 00:14:24.160
This basically existing implementation of the agent and agent came up with a new implementation where we see the v2 candidate.

s112
00:14:24.160 --> 00:14:30.959
We are still doing on format compliance and accuracy in the same way as we did before but we improved on like userf facing language.

s113
00:14:30.959 --> 00:14:34.959
So we see a positive delta which sounds like a good change now.

s114
00:14:34.959 --> 00:14:38.880
Um so uh agent reasons about it as this.

s115
00:14:38.880 --> 00:14:42.560
I mean this is a bit I mean internal application it's a bit yolo.

s116
00:14:42.560 --> 00:14:50.639
So we are like okay we are good of of just merging this and then acting again on the next change because this agent never publishes anything to actual production.

s117
00:14:50.639 --> 00:14:52.880
It just raises PS on a repo.

s118
00:14:52.880 --> 00:14:58.160
So it's okay if it kind of like auto improves and then we just provide feedback again to whatever the next increment is.

s119
00:14:58.160 --> 00:15:09.839
So um here we use our like prompt management logic so that the agent can just basically feature release this new um implementation and uh now it just summarizes basically

s120
00:15:09.839 --> 00:15:12.240
um how how we acted on this.

s121
00:15:12.240 --> 00:15:24.160
So high level um this is what we have seen like most of our best users do already uh when they when they build agents like collect production signals uh track them alongside like detailed execution traces

s122
00:15:24.160 --> 00:15:33.600
and then run agents like on a loop uh usually like a crown job like every day every week whenever you have like a batch of basically interesting user data again to then act on it.

s123
00:15:33.600 --> 00:15:44.000
Um either it's just user feedback how how it's happens in this case or internal labeling from like a production system but it could also be you ask people to annotate data like in app for example in langu we also allow like for annotations

s124
00:15:44.000 --> 00:15:58.079
to feed into that uh into that queue and yeah that's very high level what we've seen happy to give you like a more deep dive um demo at our at our booth or after this talk but just wanted to to wrap it up very concisely with how we see this this space evolves

s125
00:15:58.079 --> 00:16:08.639
and um I think what's what's interesting here is it drives the need for like a very scalable data system because you'll want to have agents really loop on this data and like produce lots of queries.

s126
00:16:08.639 --> 00:16:14.399
This we see how langu was historically very right intensive to like um ingest evils and traces.

s127
00:16:14.399 --> 00:16:19.519
Now it's way more read intensive because agents can like like chew through so much more data.

s128
00:16:19.519 --> 00:16:31.360
interesting number one and two you want to really like own the data layer because like now the traces of like for example a year ago are interesting context and you don't want to want them to be locked up in like like a more commercial system or something where

s129
00:16:31.360 --> 00:16:41.440
uh you need to sample data or retain them for a long short time period but we want to like retain them for a long time period and not sample them and yeah this really drives the need for like a scalable cheap solution

s130
00:16:41.440 --> 00:16:49.759
and we've worked with many customers on this so happy to talk one about uh oneonone about this if you're interested thank you so much for your