Why More Context Makes Your Agent Dumber and What to Do About It

Talk transcript

1. 0:00

   [upbeat music] I'm Nupur.

2. 0:15

   I work with Qodo. Uh, at Qodo, we do agentic reviews. Uh, I have a background in DevSecOps, so I'm coming from an industry where everything was deterministic. The pipelines, they run, they crash.

3. 0:29

   If they crash, we fix them. Uh, to a place where we are doing agents, where nothing is deterministic. So, in my last few years, I have learned where and how agents fail, what are the learnings, and today I'll be sharing some of my learnings with you.

4. 0:52

   So, um, if you see the evolution of agents, it started with static prompts where it was a four K context window, and we tried to put whatever was important or whatever we deemed important, and the AI models will process it and provide you with the results, right?

5. 1:12

   When we started with that, that means that it was on us to tell LLMs what do they should, uh, they should look into. That means if we provide wrong inputs, we might not get proper results.

6. 1:24

   And then we thought maybe if the context window grows, if the context size grows, we can do better, we can have more inputs. And, uh, we started with agentic workflows.

7. 1:34

   So we created an agent. We get that... get them tools like search tool to go into search into documents and do something, uh, as a command. Then again, look into the search and do something, which again created kind of a loop where the tool does not know where to stop.

8. 1:51

   It thinks like, "I need more inputs." Again, going back and back. It's a loop.

9. 1:58

   To improvise on that, uh, nowadays multi-agents is becoming more popular. Create multi-agents, do a lot of stuff together. When we see it like that, we have a lot of agents working for you.

10. 2:11

    So a security agent trying to figure security concerns, a review agent trying to review the tool, a coding agent trying to fix things. Now, again, the more the tools, the more issues you have.

11. 2:24

    Not every agent understands, and they have clash in their understandings where you don't get into the results.

12. 2:34

    So, what do we learn from here? What, uh, we see is context is not a problem. Day by day, the models are coming where you can dump a lot of context, a lot of data.

13. 2:47

    But does that make sure that the results you are getting is smart enough to give you everything or smart enough to decide what's important? If you see the current LLM models, we see a pattern where it takes the f- initial inputs you provide, it takes the last inputs, but the in-between context is basically removed.

14. 3:11

    So they don't focus on the in-between context. Agents look at the starting point, end point, and try to provide you the results. This is like a U-curve, where some of the things from the start, some of the things from the end make sense, but whatever you are providing in between that, that is not taken up.

15. 3:29

    Yeah.

16. 3:30

    How do you know this?

17. 3:32

    This is something which we are working on, and we are actually benchmarking things. So when we create agents, we try to see this is the context we provide to the agents.

18. 3:40

    Does it take this context into effect and also give us the results? So, we are working with multi-agentic architecture, where for each of the tasks we do, do for code reviews, we give the task to an agent and say, "Okay, give us the result."

19. 3:56

    Now, every time we... for example, code reviews, we try to see can we give all the context? Can we give the whole code base, for example, and see if we can get the results.

20. 4:05

    But we see that, that whenever we start working with that, the initial prompt or the initial goal which we start with, that is in focus. If we give something at the end as an input, that is in focus.

21. 4:16

    But all between context, like I have Jira, I have MCPs, can you look into that? The LLMs try to get rid of those things and push them to give...

22. 4:26

    make sense by themself. So, to s- have this or to make a way out of this, how we deal with is creating strategic solution for context optimization.

23. 4:42

    Rather than dumping everything down to the models and asking them to be smart enough to find out what is more important, uh, we usually try to see, okay, what we can do to make it a better context for the model.

24. 4:55

    There are s- lots of solutions in place if you see currently, and context engine is a buzzword. Like, everybody wants to create context engine and everybody wants to, uh, provide that.

25. 5:07

    But context engine is like a bouncer, right? So your high-speed car is going, and it acts as a bouncer and tells you, "This is more important." Now, if you have a large messy code base, it makes sense to create a context engine because it creates a search pattern, it create a ranking logic, so that whenever you ask

26. 5:28

    for a task, it looks for those rankings and say, "This is more important for you. Take it and work with it." The problem is the indexing part takes moderate effort, but the scaling is a challenge.

27. 5:41

    Like, if you start talking about six hundred repositories or seven hundred repositories, the mapping and the indexing starts to slow down, and it becomes, again, unpredictable to find or create a context engine if you are not actually into making context engine only.

28. 5:59

    There are- Lots of areas where agent can get more context instead of investing highly on context engine. Hierarchical summarization, where instead of creating or going through everything, a summary is created for each file and folder, so that when the agents try to find, they can try to read the summary and see if that is more important to

29. 6:22

    us or not, can be a good one. The only thing is that you need a lot of LLM processing. So every time a file is created or changed, some of the agents need to go and create a mapping for that.

30. 6:34

    So it's a high upfront, um, context on, uh, LLM processing that is needed. Another way is knowledge graph. Now, knowledge graph is complex, but it works wonder when you have logical dependencies.

31. 6:49

    For example, you have one file which impacts another file, which again impacts other file. You can create a graph DB hosting. It is the initial input needed by the developer is very high.

32. 7:02

    It takes a lot of time to create that. But if you have complex logics or you have mul- dependencies on multiple repos, that works wonder.

33. 7:11

    For me, I think for most of the task, if you're not a product company, but if you're building agents for yourself or your processes, iterative retrieval works really s- good.

34. 7:22

    Because instead of even creating a summary, it creates kind of an index. So it's like a library card, which you give to your agents and say, "This is the topic.

35. 7:31

    If that is relatable to you, you can look deep into the code, uh, and, uh, look for the results." Again, it has quite, uh, cost impacts, but, uh, you do not have to invest a lot of energy.

36. 7:45

    The, uh, input re- uh, required by the developers to provide to the LLMs is low, and it provides better results. There is also option of self-correction, where you ask, uh, the LLMs to do something, and there is a critic node which looks and say if that is relevant to your initial goal or not.

37. 8:06

    In that case, if the context is lost, uh, you can again ask the agent to do it again, retry it again, because the critic node said this is not the right way.

38. 8:16

    It takes a little bit more time because it ta- adds a latency of running the agents and again, but it does not require a lot of input initially from the developers to create something.

39. 8:31

    Another challenge, uh, which I have seen people getting into when they create, uh, these, um, agents is the Orchestration paradox. Now, what it does is that now LLMs are becoming more and more smart.

40. 8:47

    So when you give them the task, they think like, "Okay, I should use this tool. Uh, maybe I can do better. I should research more on what should I use."

41. 8:54

    It goes into a loop that instead of actually s- looking into solve the problem, they look for the method to solve the problem. They hop on from one method to one...

42. 9:04

    another method, and most of the API tokens are wasted on finding a way to do it rather than doing it. So you will just go on the research mode.

43. 9:15

    For example, if you use Opus latest and greatest, they will try to see what is the best method to do it and challenging themself again and again, maybe not this, another way, another way, and it just goes into a loop of doing s- trying to do something rather than doing something.

44. 9:34

    To s- resolve this, uh, we worked with 80/20 hybrid approach. I think this is one of the most interesting, uh, outcomes I have seen or the way to en- resolve this infinite loop.

45. 9:49

    What our teams are doing is giving the latest and the greatest models or giving the agents power to research 80% of the time. So you give them the goal and say, "Okay, try to do whatever you can."

46. 10:02

    But the 20% of the task where you need final validation, you want summarization, that are not something which is free-flowing. That is more restricted. Those are more hard gates.

47. 10:14

    For example, if I get X results, I want Y. It's more deterministic so that the research which is coming from the 80% can be lowered down. Now, when you see, you can always say that the 80% tool can still go on and go into infinite loops.

48. 10:32

    We have mechanisms to work on that. For some organization do, they c- they do counter mechanism, where after four or five counters, you have to work with whatever was the last results.

49. 10:45

    For some of them, they have timeout counters, and after five minutes, whatever is the last tool or whatever is the last decision, you work with that and then go back if the results are not good.

50. 10:55

    But you can restrict that 80%. But in short, if you are using anything like discovery or you're trying to see which tool to use, you're trying to plan, those 80%, uh, research models are really good.

51. 11:10

    But if you are again trying to create a summarization, you're trying to see, okay, this is the research I have got, now I have to make a result out of it, the 20% works really well.

52. 11:21

    Now, for 80%, usually you use, uh, high reasoning models, latest and the greatest, but you don't need a, a high reasoning model for the 20% because those 20% things are doing deterministic.

53. 11:34

    They are-- You are actually telling them what is needed. For example, the critic node which we talked about. They don't need to research. They don't need to find out what is the best thing to do.

54. 11:44

    They just need to see what was your goal, what was the result you are trying to achieve, and, uh, how to provide or how to summarize that for that.

55. 11:52

    Also, things like if you think about what would be the next possible action. I have this result. What should I go and look for? That are things can be done by the 80% uh, dynamic models.

56. 12:04

    Whereas I have all the results from the 80% models, but what is the proper way or what is the proper, uh, results the g- uh, the user is looking into?

57. 12:15

    Those kind of precisions can be taken by the 20% model.

58. 12:22

    Finally, uh, this is again an interesting failure which we have seen, where as the context grows, teams think, "Okay, we can do everything with one agent because the context window is quite great, right?

59. 12:35

    We can put everything. We can ask an agent to, uh, do, uh, the testing part. We can ask the da- agent to do review part. We can ask the agent all the kind of things because the context is same, and they can provide us the results."

60. 12:48

    That make sure that the... when the agent is going forward, it get overwhelmed with the inputs. And again, it tries to start losing, uh, what was the original task.

61. 12:59

    So maybe you give four tasks to the agent and somewhere down the line, it focuses on two tasks, so you get great results for the two tasks, but the other two just, just get lost in the middle.

62. 13:10

    For those particular purpose, we have something called mixture of agents, and that's, that's where you hear a lot of buzz about multiple agents or multi-agent tech architecture, where you create instead of one big agent, we create issue expert agents.

63. 13:28

    We create small, small agents which are doing great in a, a specific task which they have provided. Now, to build on top of that, each of the agent come up with their own interesting ideas or results.

64. 13:42

    How to make sure those results combine and make sense together? Because, for example, I am trying to search for a vacation. I give an agent to find the best hotel, another agent to find the best location, another agent try to find best flights.

65. 13:59

    But all three of them gives me different result. The, uh, hotel is in Greece. Uh, the flight is from Amsterdam to, um, maybe Portugal, and everything just doesn't make sense, right?

66. 14:12

    So for that particular purpose, there's a concept called a judge agent. What it does it, it tries to get all the results and see if they can make sense together.

67. 14:22

    So now you are doing all the greatest things from different agent, getting the best results from their part. But a judge agent help us to combine these and make one sense out of it instead of getting so many things which doesn't make sense together.

68. 14:41

    Something similar is implemented by us. So this is our architecture, uh, Qodo's architecture, where for the code reviews, we are using the same, uh, formula. So as part of a PR review, we have a context collector which actually goes and collect context from the PRs.

69. 15:01

    It go collect context from the context engine. It, uh, co-collect context from the tools. But then it does not start working and, uh, giving you the reviews. It actually bifurcates all the context it has provided and pass it on to different agents.

70. 15:17

    Now, what these agents do is basically specializing in what they are supposed to. For example, there will be a security agent trying to find security flaws. There might be a agent try to code.

71. 15:29

    There might, uh, code differences. There might be an agent trying to find the Jira issues. Once all these ir- agents give us back, a judge agent actually looks for the results and say, "Okay, these are interesting enough, but is it relevant to you?"

72. 15:44

    We can again go back in the context engine, look into the PRs and see out of the 10 things which is provided by you, how many of them actually make sense for your thing.

73. 15:54

    So again, refining the results, uh, to make sense to you.

74. 16:01

    Yeah, I think that was it from my side. Uh. [clapping] [laughs]

75. 16:08

    Any questions?

76. 16:10

    Yes. Um, in practice-

77. 16:13

    Mm-hmm

78. 16:13

    ... how do you let the swarm communicate with each other?

79. 16:17

    Uh, you are talking about the agents?

80. 16:20

    So you have agent A and agent B and the judge. I can imagine they write to a file system, or do you have some kind of tool proprietary or?

81. 16:28

    Uh, we ch- we use, uh, LangChain at the bottom, uh, and that is being used to communicate and build infrastructure for, uh, different agents.

82. 16:38

    Do you know what LangChain uses for that? Like, just collects the responses and then shoves it back into the prompt of the next agent?

83. 16:47

    Yes. Yes. Yes. Yes. That- that's what. So what we do is we try to cr- uh, get the results and create a prompt for the next agent. And if it's multiple things, again, there is an agent just to collect the results and create a better prompt, uh, which is refined for the next agent.

84. 17:05

    Okay.

85. 17:06

    Have you thought about a calibration step for each agent?

86. 17:10

    When you say the calibration, can you t- tell me more-

87. 17:12

    The calibration, right? So you... When you do a code review with an agent, right? You need at least, what I heard today multiple times is doing some kind of calibration, right?

88. 17:25

    That you actually tell that what is good and what is bad.

89. 17:28

    Yes. So when you say it like that, uh, and let me know if that makes sense for your question or not. We do calibration in a form. We check what we have as a context.

90. 17:40

    So for example, uh, when we get the code reviews, LLM does not know what is important for you or how do you work. So for example, an LLM when they gets input, they get input from healthcare industry, they get from retail industry, they get from finance industry.

91. 17:57

    And all of them can use same Java framework in different ways. Uh, different, uh, things are important for them. The rest of them doesn't make sense. So what we do is we give you two different options to tell the agents how to perform or what to work on.

92. 18:15

    One, uh, part is we give them the PR history. So we index all your PR and see when was the last time something like this was identified and compare the current sub-version, if that is-

93. 18:30

    Yes, yes, yes, yes. We do the... So the changes we, uh, you make to the code, we look into if we can see something similar in the past. That again is, uh, transferred to the context twice.

94. 18:42

    First is when we are actually giving context to the sub-agents to find things for you, and another time to the judge agent, so that when I get fifteen different recommendations for your code review, my judge agent can look into what was there before, how did your reviewer commented, how your developers commented, and based upon that decide if

95. 19:03

    that is worth providing to your developers or not. And the other part is-

96. 19:07

    And this happens for every agent?

97. 19:09

    For every agent, yes.

98. 19:10

    Okay. And, and if I understood you right, you don't share the context between the agents, right? So you have the, the co- a specific context for every agent.

99. 19:22

    Yes. We are trying to resolve that new part that instead of dumping everything to the LLM, we, uh, take the part which is more important. We use a context engine for that.

100. 19:33

     We take the part which is more important and only provide that particular part to that particular agent.

101. 19:39

     Sorry, one more follow-up.

102. 19:40

     Yeah, sure.

103. 19:41

     Um, but I... But then for me it's not clear how you bridge the gap, right? Let's say you have an, uh, code quality agent, and you have a, I don't know, uh, um, what else agent, um, uh, uh, framework-specific coding review agent, right?

104. 20:03

     And then you basically, you only, as I understood-

105. 20:08

     Mm-hmm

106. 20:09

     ... um, you only share the specific information to each agent, right? And then basically each agent runs atomic, autonomously-

107. 20:18

     Yeah

108. 20:18

     ... right? Um, and doesn't have the full picture, right? And at least when, let's say as a human, right, and if you do code reviews, it's always good if you have a full prospect, right?

109. 20:33

     I, I would say, like, that kind of methodology would... it works for simple things like does it use linting? Does it use, I don't know, i- are tests implemented?

110. 20:43

     But when you think about, or at least I think when you think about the overall architecture, for example, to make architecture decisions, um, that covers security because everything is a balance and a, and a-

111. 20:57

     Yeah

112. 20:57

     ... and you have to weigh that somehow. So how, how do you solve that then?

113. 21:02

     Yes. So I think if you look into the older v- version of code reviews, you should ha- you used to have a senior engineer who knows your code, who knows what kind of packages you are using, and they can, uh, comment on if the developer has done something, uh, similar to what you're used to or did something,

114. 21:20

     uh, totally weird, right? Then you used to have a, a security person who used to see if you are providing all kinds of security. You are not hard coding your APIs, or you are not, uh, putting any SQL injections.

115. 21:33

     So all those kind of things the security experts know. On the other hand, if you are working with ISO compliance or SOC 2 compliance, there might be an auditor who might ask the, uh, team lead or the senior engineer, "Is your code, uh, being, um, you know, logged?

116. 21:50

     I-Is it logging the changes?" and so on. So previously as well, there were lots of people having specialized knowledge looking into those kind of areas specifically. Now, when the context is provided, it's always like, these are my security concerns which I always have to look into.

117. 22:05

     These are my architectural concerns. An architect might look from the architecture per- perspective. We can do that something similar, uh, with the agents as well because, for example, architecture security concerns, we have a web portal where architects can provide their guidelines for, uh, co- compliance people can provide their guidelines, and an agent can look into all those

118. 22:25

     guidelines and say, "Is it validated or not?" So if you see the

119. 22:31

     initial, uh, picture, the context collector knows everything, and then it provides relevant context to the agents.

120. 22:39

     So you, you enforce basically your customer to upload that, these kind of documents, right? Is that a kind of a requirement then? Or-

121. 22:48

     It-

122. 22:49

     Because, I mean, the system will have completely, let's say, different result, right, if you don't share them.

123. 22:55

     Exactly. So that's something which depends upon organization. There are some people who say, "We don't work with any rules or regulations, so just give me out of the box."

124. 23:04

     That also means that don't expect the agents to find something very specific to your working until and unless we have certain PR history. Because then again, the PR history kicks in and tells you what is relevant even then when you don't provide.

125. 23:19

     I'm not sure if the, if the PR history is really the best source, right? It-- I think it-

126. 23:26

     It can be one. It can be one of... It can-

127. 23:29

     Yeah.

128. 23:29

     So that's why there are various, uh, sources, right? So it's PR history, it's your reso- it's your... [phone message alert]

129. 23:35

     And somebody's cooking food at my home. [laughs] Yeah. But it, it can be, uh, one of the source. And that's why-

130. 23:42

     Yeah. And my question is like... Yeah, I mean, of course it can, it can be, right?

131. 23:45

     Yeah.

132. 23:46

     But at the end you need to decide, let's say also with jury, right, how much you weight, right? So like the new documents for the, let's say, the engineering principles, architecture principles, and compare, compare them to, let's say, your, your merge request-

133. 24:04

     Yeah, yeah, yeah

134. 24:05

     ... right away, right? But they can be completely out of the balance, right?

135. 24:09

     Um, it depends. It depends. So again, uh, I think it's, if you look into from one perspective, it's difficult to decide. But if you're getting the context from many angles, for example, PR history, that's one part.

136. 24:22

     But when you do the compliance and you tell in the compliance portal, this is really important. So we have various, uh, segments of it. It's an error, it's an, uh, a recommendation.

137. 24:32

     All those kind of things adds weight to a feedback to say if it's good or not. And every time your developer expe- uh, accepts a suggestion, it gets more weighted for the next one.

138. 24:44

     If it does not, uh, uh, accept the suggestion, it gets a less weight. So it's all about indexing and making sure those weights are managed somewhere.

139. 24:53

     How do you weight them?

140. 24:54

     Uh, two ways. One is, uh, by when you, when you give your recommendations, does your developer actually accept it or not? We, uh, index that.

141. 25:03

     Right.

142. 25:03

     Another way is, uh, from the past PRs, we try to find out similar issues identified and if your developer actually implemented them. So for example, some people are used to hard coding their, uh, API keys, and I literally had a tough, um, argument with the developer.

143. 25:21

     But this is how we do it. No, this should, this should not be the way.

144. 25:24

     And then-

145. 25:24

     Yeah, but-

146. 25:25

     That's, uh, and I think [phone message alert] it nicely matches also what I meant, right? So if you look in the, in, in the history-

147. 25:32

     Yeah

148. 25:32

     ... that happened doesn't mean it's, it's good.

149. 25:34

     It's g- Yeah. A-and, and that's the way where the system tries to tell you, "This is not good. This is not good." And then it's up to you and-

150. 25:41

     But only if you provide the guidance, right? Or-

151. 25:44

     No. It, uh, so there is something called bug fixes, and there is something called rules. So if you provide them as a rule, it will get, uh, highlighted, doesn't matter if you want it or not.

152. 25:55

     And then there are bugs where agent try to tell you there's something wrong, and if the reviewer also agrees with it and did not implement, uh, 10 times, the reviewer might get it less weighted and give you.

153. 26:06

     Okay. Yeah. Interesting. Cool. Thank you so much.

154. 26:11

     Thank you. [outro music]