How to Generate Mergeable Code with a Context Engine — Peter Werry, Unblocked

Before agents, you were the context layer

00:12What we do at Unblocks is we build a context engine. I just want to do a quick sound check at the back to make sure everyone can hear me fine. Can you guys Yeah, we're good. Awesome. So at a high level, a context engine delivers organizational context to both your human workers and now increasingly your agents.

00:27Okay. So why why is that important? Before we go too deep on the mechanics of how a context engine works, I just want to talk briefly about the problem. So, what we're going to do is we're going to hop into our time machines and we're going to travel back to the before times uh before agents and uh discuss a little bit about what we used to do as humans uh before agents came into the picture.

00:58And so for years um you were the context layer. You um had to go and do things like this. you had to find things you were looking for, trolled all over different data sources, different discussions taking place. Um, and then through the codebase of course to try to build up tribal knowledge and, uh, throughout time as your code base progressed, um, you'd be, you know, fighting incidents and things like that.

01:29And your organization over time builds up battle scars um from all these all these different things building code uh documenting architecture and and dealing with outages and things like that. But now um we have a new problem because uh as we introduce agents to the picture

An agent is a new employee who resets every task

01:49um they suffer from all of these challenges except for one thing. Agents are like new employees. they reset their knowledge every time you start a new task. Okay? And so you can think of an agent like an expert software engineer um who's a new employee on boarding for the first time.

02:11Every time they have to rediscover your code base, how your organization builds tests um and how they deploy software with each and every task. Uh can I just uh put a put a show of hands for everyone that's seen this slide before by Vim?

The maturity curve: autocomplete to software factories

02:30So this is kind of like u this is a good way to view where people are on what we call like the AI maturity curve. Um starting at the the far left uh this is kind of representative of autocomplete back in the GBT35 days. You know remember co-pilot and things like that.

02:47Um and then you know kind of move on to using cursor. Um and then from there you're you're think you're talking about how you can start to solve the context problem. So some people are building organizational wikis. Just smile if if this is kind of um bringing up memories for you.

03:04Um and then you know all these things are great except that uh how do you give agents access to this and what are the compounding problems that the scaling problems as you move forward well if you give MCP and skills to your agents um to teach them how to navigate and build context and that's kind of where uh people are today most people they're at the sort of stage four to five level okay and uh they understand that context is the bottleneck and they're trying to build solutions to solve it for their engineering teams.

03:41So looking ahead uh to all the way to eight with software factories. This is kind of where the puck is going. I'm not sure if if folks were at the keynote this morning, but um it's it's all about like delivery of context and unknown and unknowns.

03:55And this becomes increasingly important as people start thinking about full automation of agents. they just can't operate without organizational context. They get lost. So, you know, like that's the real problem. Access to information doesn't equal understanding.

04:11Um I I know that

Satisfaction of search, borrowed from radiology

04:19folks are probably familiar with claude.md um and and uh and wiki layouts and all these things. If you attach a wiki, it still doesn't tell the agent where the information is that it needs. It can search for things in the wiki, but then what happens is it'll suffer from something that uh radiologists uh call satisfaction of search.

04:42So, this is a term in radiology where you look at an X-ray and you're trying to find a region um that might be an indicator for cancer. Okay? And you discover like one indicator and if you stop there uh you might miss other important indicators that might you know lead to diagnosis of even more uh issues.

05:06So this is what happens with agents. They don't they they find something that they they think is correct and then they stop. Um the the other thing about agents is that they don't distill understanding. They can look around, they can find information, but they they don't understand how all the pieces fit together because without doing that leg work ahead of time.

05:29Um, they don't understand how, you know, your dependencies interact with each other and how your architecture and sort of future planning is going to scope the work that it does next. And so some some people will then ask, well, what if we just take the entire codebase and all of our architecture documents and just slam it into the context window.

05:45Um, and then yes, maybe like your agents will reason about everything all at once. And in practice, that that of course doesn't work. Um, not just because you've got way more organizational context than can fit into a context window, even one that's a million tokens in size.

06:02Um, but it it it causes the agent to get distracted. When you're working on a task, you want task specific flow. Um, and so your agents will get distracted easily if you give them things that cause them to look this way in that way. Um, and it'll just waste tokens and time.

06:19So, in this morning's keynote, um, Tariq from Claude Code mentioned unknown unknowns. I just want to uh harp on that phrase again. And it can be phrased a different way, which is finding the things that really matter.

The iceberg: intent, conventions, past decisions

06:37And so this is what your agent can see at the top of the iceberg. They can see the code and they can operate on the code. What they don't see are things like the actual intent, the team conventions, past decisions, things that you've discussed in Slack, for example, uh architecture rationale, and so on.

06:59And that's why your agents need a context engine to get real work done. So, I'm going to now uh attempt a live demo. And hopefully the demo gods are kind. Um, so I want to pop back up conceptually. Oops, I think I'm on the wrong tab. We'll get to that one in a sec.

07:20So for now, sorry about that. And here we are.

Demo: asking about a component, and showing the work

07:29So I'm going to ask a question as if I'm a, you know, I'm a human and I want to get some information about my codebase. And, you know, the human layer hasn't gone away. We talk about agents and their need for context, but um humans are still asking questions about the codebase and we need that level of understanding because ultimately the accountability stops with us.

07:46When you hit merge on a PR, you need to understand what it's doing um and you need to understand how the architecture works. So this question I asked here um is about an internal component of our system called the source mark engine and you can see that it uh is able to articulate it fairly well.

08:02um understands the architecture. This this diagram here is uh is generated. So it this diagram doesn't exist. Um it just figures it out based on the um the way the code operates today and then some proposals for future architecture. And then uh what's really important is that you show your work.

08:27This is a trust building thing more than anything, but it allows people to see if um if the answer is maybe not entirely correct, then you can in look into the uh the knowledge base that you have and make corrections. Increasingly agents are doing this for you.

08:49So now um what I want to show you is another place where humans spend their time which is in Slack and this is where a lot of the decisions get made of course. So I can do something like this. And uh unblocked will sit and kind of listen for things that are things that can chime in on when it provides a high degree of Oh, sorry.

09:04We went to the wrong You guys can't see that. Thank you, Claire. Oh, come on down. Let's see if I can bring it up. There we go. Perfect. So I can ask questions like this in unblocked and if it thinks it can chime in on the answer then it will chime in.

09:24Otherwise I can just um address unblocked directly and ask the same question and when it thinks that it has an answer to give then it will give an answer and so we can get um quite a bit of interesting content there from unblocked. Thank you.

09:43Unblocked. I'm going to switch up and show you the the really interesting thing which is the agents. Okay. So, um

The same plan with and without a context engine

10:03in in that question, the source mark engine, I'm not sure if people picked up, but there was a little thing at the bottom there that said, you know, there's some optimization opportunities. Um so what I did here is I went into claw code and I asked it um without using unblocked to um uh generate uh a plan to optimize the source mark calculator and it did that and it happily went and you know searched through the code and and tried to figure out how the algorithm works and so on.

10:27Um and it it reached a conclusion that's great you know it does a pretty good job um but you know it maybe could do a little bit better. So, I asked that question again uh using unblock this time and it it really kind of nails the the nuances because it picks up on the the uh PRs that we um where we discussed future possibilities for improvement.

10:48um some Slack conversations that we had and uh of course you know notion and architecture documents and it shows its work and this is really important because um all of these things here the sources come back to Claude and then Claude knows exactly where to jump to next if it needs to elaborate on that context.

11:13And so I just want to show you what the impact of that is. So if I um Whoops. Thank you. If I pull up usage here, you can see that with unblocked, uh, the total cost was, you know, subd dollar to create the plan. Uh, took about a minute. Ignore the wall clock time because I've had this open for about an hour.

11:31But, um, it's about a minute. And then if I look at um the usage without unblocked, you can see that it's about 2 minutes. And and and it costs more to generate all that context. Now, the reason that happens is because it has to do more work.

11:49It has to look around. has to discover things. Um, and this compounds, not only does it have to do more work to discover things, it doesn't discover the right things. So, when you get further down in your execution, it may be operating on the wrong plan or the wrong assumptions.

12:06And then you have to go back and you have to loop over and over again. So, the real value of a context engine is not like the upfront cost on these short tasks. It's the compounding effect. Um the the other Tariq from Sonar mentioned this in the keynote this morning and it's true like the loops compound and you have to be like um uh efficient the entire way through with your context.

12:30I'm just going to jump back to Safari and I'm going to point out um some really interesting things. So we also have a a code review agent.

A review agent that boosts what senior engineers said

12:44And when we say um you know organizational context, we're talking about more than just the underlying data. Uh we're talking about real intelligence. So what unblock does is it looks at um not like it looks at pull request data and there are other data sources for this and it generates a series of best practices that help align agents to your codebase.

13:08But we thought that this would be really helpful to surface for the review agent as well. So what you can see here is um it unblock chimed in and then Richie here said, "Oh, that's cool. That's something I would say." And that's because that actually was something he said.

13:22So it surfaced the uh the previous comments. Richie's one of the senior engineers and we use the sort of seniority or expertise as a signal um to boost uh comments that are important. Okay. So another uh interesting interaction by Richie, he uh discovered that the number of code review issues that were being

Debugging a drop in flagged issues, back to Slack

13:50surfaced dropped uh precipitously and he was debugging it with unblocked. Um he got all the way to the bottom and realized what roughly what the problem was and then asked unblocked to fix it. Now this this is something that we have internally um you know that we're experimenting with.

14:02Um, so unblocked uh can run as an agent in the cloud. Um, but what's really cool about this is that it has all your organizational context at its fingertips and the results are are pretty magical. So it can do things like generate this PR um, and then what you'll see here is that not only does it generate the fix, it also is able to relate it to the all the conversations that were happening.

14:26So this PR was created because and you read that context thing. It's mind-blowing. After this PR, we switched to uh Claude 48 and it dropped a ton in issues because of the behavior is quite a bit different. So then it said Richie directly correlated the drop.

14:48Now what's this thing here? Let's click on it. It is a Slack conversation. So, it found the Slack conversation, correlated all of that, you know, past history back again, and then we ended up with a with a final PR.

Open source: query engine and social graph

15:08So, um, I'm going to I've got only a few minutes left. I'm just going to close this out really quickly. We have a uh a couple of open- source projects that are kind of interesting if people want to play with them. One is the document query engine.

15:16That was, uh, something that I talked about on Monday in my workshop. Um, I may uh talk about it again tomorrow, but I just want to give folks a sense of what this thing does. Um, whoops. If you want to play with it, it's open source, so you can just download it and have it go.

15:32It basically runs over your um uh GitHub repository, ingests uh your your historical pull requests, and then uh synthesizes a schema based on the documents that it can sample. Um and then from there you can issue any kind of queries that you like and get all kinds of insights out of it through the agent chat.

15:57You can ask all kinds of questions. Um and then lastly the engineering social graph. So this is the thing that I was talking about earlier that helps us pin down expertise and team relationships. Um so what you can see here is this sort of like the rough breakdown of our team structure at Unblocked.

16:12As you can see we're a fairly small team. Um and so we've got these um uh these clusters of people and how they relate to each other indicates the kind of um review relationships that they have. So these are you know these lines show like we review each other's code.

16:35Um we can then cluster that and generate team labels for that or show the coverage across your codebase. This is really cool. you can see kind of where the holes are, where you might be lacking expert coverage. Um, and that's exactly what we use within the context engine itself.

16:54All right, one last thing we have uh for those that want a taste of what a context engine can do but don't want to sign up for unblocked right away. Um you can use uh something that we call the context engine simulator which will basically build up a context behind the scenes on a per task basis and then use that context uh to to drive the task.

17:18It'll do it with context and without context so that you can see what the differences might be. This is a QR code for that if you want to just take a quick snap. Awesome. And I'll just land on a quote from one of our customers. 50% fewer tokens, faster triage, better answers.

17:45And that's exactly what a context engine can do. One last shout out um before we end. My colleague Brandon is giving a talk in 10 minutes uh at room 2020. um he's going to speak to in a lot more detail about some of the higher level things that context engines can do.

18:08I'm going to run over there right after this and I think all of you should follow me. Awesome. Oh, and don't forget to get a coconut.