AI Engineer World's Fair 2026

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

Read the talk

How to Generate Mergeable Code with a Context Engine

Peter Werry explains why coding agents need selected organizational knowledge—not merely repository access—and demonstrates how Unblocked connects code, decisions, review history, and expertise to improve planning and debugging.

From a talk by Peter Werry

At a glance

Ideas worth remembering

  • Coding agents repeatedly rediscover repository structure, testing practices, deployment conventions, and organizational history because each task begins without the accumulated knowledge of an experienced teammate.

  • Search access alone can produce “satisfaction of search”: the agent finds one plausible result and stops before discovering evidence that would change its plan.

  • Useful context is selected for the task and connects code with intent, conventions, past decisions, discussions, and architecture rationale. Dumping everything into a large context window can exceed capacity and distract the agent.

  • Sources make synthesized answers inspectable by humans and give coding agents direct references for deeper investigation.

  • The planning demo reports about one minute and a sub-dollar cost with Unblocked, versus about two minutes and higher cost without it. The more consequential proposed benefit is avoiding downstream loops caused by incomplete discovery and wrong assumptions.

  • Pull-request history can encode team-specific review practices, while expertise and review relationships can prioritize guidance and expose parts of the codebase with thin expert coverage.

The context layer agents keep losing

A context engine delivers organizational knowledge to both people and agents. Before coding agents, engineers assembled that knowledge themselves: they searched discussions and documents, read the codebase, documented architecture, and learned from incidents and outages. Those experiences produced the “battle scars” that explain why a system works as it does—not just what the current code says.

Shows the AI maturity curve as Werry locates many teams around the stages where context becomes the bottleneck.
Shows the AI maturity curve as Werry locates many teams around the stages where context becomes the bottleneck.

An agent starts from a harder position. Werry compares it to an expert software engineer joining the company for the first time on every task. General programming ability does not supply knowledge of this repository, its testing conventions, or its deployment process. Once the task ends, the newly discovered context effectively disappears, so the next task begins with another onboarding exercise.

He places this problem on an AI maturity curve that moves from autocomplete and editor assistance through organizational wikis, MCP-connected tools, skills, and eventually software factories. He puts many teams around stages four to five, where they recognize context as a bottleneck and are trying to make internal knowledge available to agents. More autonomous systems raise the stakes: an agent working without close supervision must find relevant facts that it does not already know to request.

0:120:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

Why search access and a huge prompt are insufficient

Attaching a wiki gives an agent somewhere to search, but it does not tell the agent what matters or when the search is complete. Werry borrows “satisfaction of search” from radiology: after finding one plausible indicator, a reader may stop and miss another finding that would change the diagnosis. A coding agent can make the same mistake by finding one apparently relevant page or code path and proceeding before it discovers a conflicting decision elsewhere.

Displays the key distinction that access to information does not equal understanding.
Displays the key distinction that access to information does not equal understanding.

Retrieving individual facts also leaves a synthesis problem. The agent must connect dependencies, architecture, and future plans to determine the scope of a change. A plausible document is not enough if it does not explain how those pieces constrain one another.

Putting everything into the context window fails differently. Werry argues that an organization can hold more relevant material than even a million-token window can contain, while irrelevant material distracts the agent and consumes time and tokens. The target is task-specific context: enough evidence to direct the work without filling the prompt with unrelated code and documents.

His iceberg model separates visible code from the knowledge underneath it. Agents can inspect and edit the code, but intent, team conventions, past decisions, Slack discussions, and architecture rationale may live elsewhere. Those submerged constraints are often the difference between code that merely compiles and a change that fits the organization.

How it fits togetherThe organizational-context iceberg

The implementation the agent can inspect and modify directly.

The repository exposes implementation, while many constraints that determine the right change live in organizational history.

4:114:41
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:11 · section reference included

A generated architecture explanation that shows its work

The first live demonstration returns to the human reader. People remain accountable for what they merge, so they still need to understand a pull request and the architecture it changes. The demo asks Unblocked about an internal component called the Source Mark Engine.

Shows the answer’s attached sources while Werry emphasizes that synthesized explanations should show their work.
Shows the answer’s attached sources while Werry emphasizes that synthesized explanations should show their work.

The answer includes an architecture diagram synthesized from the current code and proposals for future architecture. It was not retrieving a diagram that already existed. It inferred a representation for this question, which makes provenance especially important: a generated explanation can be useful and still be partly wrong.

The attached sources let a person inspect the knowledge behind the answer and correct it. They also give an agent concrete places to continue investigating. “Show your work” is therefore more than a trust gesture; it creates a path from synthesis back to the underlying evidence.

Werry then moves the same question-answering behavior into Slack, where many engineering decisions happen. Unblocked can respond when it judges that it has a sufficiently useful answer, or a user can address it directly. The demonstration does not explain how that confidence decision is calculated, so selective participation remains a stated behavior rather than a fully specified mechanism.

7:297:59
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:11 · section reference included

The same optimization plan, with and without organizational context

The Source Mark Engine answer identifies optimization opportunities, which become the next experiment. Werry asks Claude Code to plan an optimization of the Source Mark calculator without Unblocked. The agent searches the repository, studies the algorithm, and produces a plan that he describes as reasonably good.

Shows the usage comparison while Werry reports roughly one minute and sub-dollar cost with Unblocked versus a slower, costlier baseline.
Shows the usage comparison while Werry reports roughly one minute and sub-dollar cost with Unblocked versus a slower, costlier baseline.

The second run adds Unblocked. This time, the plan incorporates pull requests discussing future improvements, Slack conversations, Notion material, and architecture documents. The response carries its sources into Claude Code, giving the coding agent specific references to inspect if it needs more detail rather than forcing it to rediscover the relevant trail.

For this planning demonstration, Werry reports a cost below one dollar and a task time of about one minute with Unblocked. Without it, the run took about two minutes and cost more. He explicitly says to ignore the displayed wall-clock duration for the run with Unblocked because its session had remained open for about an hour. These figures describe one planning demo; they do not establish a general benchmark or show that either plan was implemented and merged.

The larger claim is about downstream rework. Baseline discovery costs time, but incomplete discovery can also put the agent on the wrong plan or wrong assumptions. Each later step then inherits the mistake, eventually forcing the agent to revisit earlier work. The initial minute saved is small compared with avoiding repeated loops across a longer task.

How it fits togetherWhy missing context compounds

Plan an optimization for the Source Mark calculator.

A discovery mistake can survive planning, contaminate execution, and force another search loop. Selected cross-source context aims to reduce that risk before execution starts.

9:5510:25
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:55 · section reference included

Turning review history and expertise into guidance

The code-review example treats organizational context as more than a collection of documents. Unblocked examines pull-request data and other sources, derives best practices associated with the codebase, and makes them available to a review agent. The intended result is guidance shaped by how this team has reviewed real changes.

Shows the generated pull request connecting the proposed fix to the earlier model-behavior regression and organizational history.
Shows the generated pull request connecting the proposed fix to the earlier model-behavior regression and organizational history.

In the demo, a senior engineer named Richie recognizes an automated comment as something he would say. Werry explains that the system had surfaced a comment Richie previously made. Seniority or expertise acts as a boosting signal, giving some historical guidance more prominence. The talk does not specify the scoring formula or explain how the system handles conflicting expert opinions.

A second example begins when Richie notices a sharp drop in the number of issues surfaced by the review agent. He investigates with Unblocked, reaches an approximate diagnosis, and asks it to fix the problem. Werry describes this cloud-agent behavior as an internal experiment rather than a generally established workflow.

The generated pull request ties its fix to earlier conversations. A change in model behavior had been associated with the drop, and Unblocked found the Slack thread where Richie had made that connection. The example demonstrates traceability from an observed regression through organizational history to a proposed fix. It does not establish the exact patch mechanics or provide a measured post-fix recovery.

12:2512:55
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:55 · section reference included

Query history, map expertise, and compare task outcomes

Werry closes with two open-source components. The document query engine runs over a GitHub repository, ingests historical pull requests, samples those documents, and synthesizes a schema. Users can then issue questions through an agent chat. The talk does not detail the generated schema or the query-validation process, but the project turns repository history into structured, queryable material rather than relying only on similarity search.

Shows the context engine simulator that runs a task with and without assembled context for direct comparison.
Shows the context engine simulator that runs a task with and without assembled context for direct comparison.

The engineering social graph extracts a different kind of context: relationships among people. Connections indicate who reviews whose code, and clusters can produce team labels or show coverage across a codebase. Thin areas reveal where expert review may be missing. Werry says the context engine uses this graph itself, making review relationships both a map of organizational expertise and a retrieval signal.

Finally, the context engine simulator builds context for one task and runs that task with and without it. The design makes the comparison visible rather than asking users to accept a general claim about context quality. No direct URL is supplied in the recording; attendees receive it through a QR code.

Werry ends with a customer report of “50% fewer tokens, faster triage, better answers.” The recording does not provide the workload, sample size, or evaluation method behind that percentage, so it should be read as a customer-specific result rather than a general guarantee. The durable lesson is narrower: context engines are most valuable when organizational knowledge changes the plan, prevents a bad assumption, or points the agent toward evidence it would not have known to seek.

15:0815:38
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:08 · section reference included

Resources

From the talk

  • The product demonstrated in the talk for connecting code, conversations, issues, documentation, and organizational knowledge.

  • An open-source workshop project that turns natural-language questions into validated MongoDB aggregation pipelines over GitHub pull-request and issue data.

Read the complete timestamped transcript
  1. 0:12

    What we do at Unblocks is we build a

  2. 0:13

    context engine. I just want to do a

  3. 0:15

    quick sound check at the back to make

  4. 0:17

    sure everyone can hear me fine. Can you

  5. 0:18

    guys Yeah, we're good. Awesome. So at a

  6. 0:22

    high level, a context engine delivers

  7. 0:25

    organizational context to both your

  8. 0:27

    human workers and now increasingly your

  9. 0:29

    agents. Okay. So why why is that

  10. 0:32

    important? Before we go too deep on the

  11. 0:34

    mechanics of how a context engine works,

  12. 0:37

    I just want to talk briefly about the

  13. 0:39

    problem.

  14. 0:42

    So, what we're going to do is we're

  15. 0:44

    going to hop into our time machines and

  16. 0:46

    we're going to travel back to the before

  17. 0:48

    times uh before agents and uh discuss a

  18. 0:52

    little bit about what we used to do as

  19. 0:53

    humans uh before agents came into the

  20. 0:56

    picture.

  21. 0:58

    And so for years um you were the context

  22. 1:02

    layer. You um had to go and do things

  23. 1:06

    like this. you had to find things you

  24. 1:09

    were looking for, trolled all over

  25. 1:12

    different data sources, different

  26. 1:14

    discussions taking place. Um, and then

  27. 1:17

    through the codebase of course to try to

  28. 1:18

    build up tribal knowledge and, uh,

  29. 1:21

    throughout time as your code base

  30. 1:24

    progressed, um, you'd be, you know,

  31. 1:27

    fighting incidents and things like that.

  32. 1:29

    And your organization over time builds

  33. 1:31

    up battle scars um from all these all

  34. 1:34

    these different things building code uh

  35. 1:37

    documenting architecture and and dealing

  36. 1:39

    with outages and things like that.

  37. 1:42

    But now um we have a new problem because

  38. 1:46

    uh as we introduce agents to the picture

  39. 1:49

    um they suffer from all of these

  40. 1:52

    challenges except for one thing. Agents

  41. 1:54

    are like new employees.

  42. 1:56

    they reset their knowledge every time

  43. 1:59

    you start a new task. Okay? And so you

  44. 2:03

    can think of an agent like an expert

  45. 2:05

    software engineer um who's a new

  46. 2:08

    employee on boarding for the first time.

  47. 2:11

    Every time they have to rediscover your

  48. 2:12

    code base, how your organization builds

  49. 2:15

    tests um and how they deploy software

  50. 2:18

    with each and every task.

  51. 2:22

    Uh can I just uh put a put a show of

  52. 2:25

    hands for everyone that's seen this

  53. 2:27

    slide before by Vim?

  54. 2:30

    So this is kind of like u this is a good

  55. 2:32

    way to view where people are on what we

  56. 2:34

    call like the AI maturity curve. Um

  57. 2:37

    starting at the the far left uh this is

  58. 2:40

    kind of representative of autocomplete

  59. 2:42

    back in the GBT35 days. You know

  60. 2:45

    remember co-pilot and things like that.

  61. 2:47

    Um and then you know kind of move on to

  62. 2:49

    using cursor. Um and then from there

  63. 2:52

    you're you're think you're talking about

  64. 2:54

    how you can start to solve the context

  65. 2:56

    problem. So some people are building

  66. 2:58

    organizational wikis. Just smile if if

  67. 3:01

    this is kind of um bringing up memories

  68. 3:04

    for you. Um and then you know all these

  69. 3:07

    things are great except that uh how do

  70. 3:11

    you give agents access to this and what

  71. 3:13

    are the compounding problems that the

  72. 3:16

    scaling problems as you move forward

  73. 3:18

    well if you give MCP and skills to your

  74. 3:20

    agents

  75. 3:22

    um to teach them how to navigate and

  76. 3:24

    build context and that's kind of where

  77. 3:28

    uh people are today most people they're

  78. 3:30

    at the sort of stage four to five level

  79. 3:33

    okay

  80. 3:35

    and uh they understand that context is

  81. 3:37

    the bottleneck and they're trying to

  82. 3:39

    build solutions to solve it for their

  83. 3:41

    engineering teams. So looking ahead uh

  84. 3:44

    to all the way to eight with software

  85. 3:47

    factories. This is kind of where the

  86. 3:48

    puck is going. I'm not sure if if folks

  87. 3:51

    were at the keynote this morning, but um

  88. 3:53

    it's it's all about like delivery of

  89. 3:55

    context and unknown and unknowns. And

  90. 3:58

    this becomes increasingly important as

  91. 4:00

    people start thinking about full

  92. 4:02

    automation of agents. they just can't

  93. 4:04

    operate without organizational context.

  94. 4:06

    They get lost.

  95. 4:11

    So, you know, like that's the real

  96. 4:13

    problem. Access to information doesn't

  97. 4:16

    equal understanding. Um I I know that

  98. 4:19

    folks are probably familiar with

  99. 4:20

    claude.md

  100. 4:23

    um and and uh and wiki layouts and all

  101. 4:26

    these things. If you attach a wiki, it

  102. 4:29

    still doesn't tell the agent where the

  103. 4:32

    information is that it needs. It can

  104. 4:35

    search for things in the wiki, but then

  105. 4:37

    what happens is it'll suffer from

  106. 4:39

    something that uh radiologists

  107. 4:42

    uh call satisfaction of search. So, this

  108. 4:44

    is a term in radiology

  109. 4:47

    where you look at an X-ray and you're

  110. 4:49

    trying to find a region um that might be

  111. 4:52

    an indicator for cancer. Okay? And you

  112. 4:55

    discover like one

  113. 4:57

    indicator and if you stop there uh you

  114. 5:01

    might miss other important indicators

  115. 5:03

    that might you know lead to diagnosis of

  116. 5:06

    even more uh issues. So this is what

  117. 5:10

    happens with agents. They don't they

  118. 5:11

    they find something that they they think

  119. 5:13

    is correct and then they stop. Um the

  120. 5:16

    the other thing about agents is that

  121. 5:17

    they don't distill understanding.

  122. 5:20

    They can look around, they can find

  123. 5:22

    information, but they they don't

  124. 5:24

    understand how all the pieces fit

  125. 5:26

    together because without doing that leg

  126. 5:29

    work ahead of time. Um, they don't

  127. 5:31

    understand how, you know, your

  128. 5:32

    dependencies interact with each other

  129. 5:34

    and how your architecture and sort of

  130. 5:36

    future planning is going to scope the

  131. 5:38

    work that it does next. And so some some

  132. 5:41

    people will then ask, well, what if we

  133. 5:43

    just take the entire codebase and all of

  134. 5:45

    our architecture documents and just slam

  135. 5:47

    it into the context window. Um, and then

  136. 5:50

    yes, maybe like your agents will reason

  137. 5:52

    about everything all at once. And in

  138. 5:54

    practice, that that of course doesn't

  139. 5:56

    work. Um, not just because you've got

  140. 5:58

    way more organizational context than can

  141. 6:00

    fit into a context window, even one

  142. 6:02

    that's a million tokens in size. Um, but

  143. 6:06

    it it it causes the agent to get

  144. 6:07

    distracted. When you're working on a

  145. 6:10

    task, you want task specific flow. Um,

  146. 6:13

    and so your agents will get distracted

  147. 6:15

    easily if you give them things that

  148. 6:17

    cause them to look this way in that way.

  149. 6:19

    Um, and it'll just waste tokens and

  150. 6:21

    time. So, in this morning's keynote, um,

  151. 6:25

    Tariq from Claude Code mentioned unknown

  152. 6:28

    unknowns. I just want to uh harp on that

  153. 6:30

    phrase again. And it can be phrased a

  154. 6:32

    different way, which is finding the

  155. 6:34

    things that really matter.

  156. 6:37

    And so this is what your agent can see

  157. 6:39

    at the top of the iceberg. They can see

  158. 6:41

    the code and they can operate on the

  159. 6:44

    code. What they don't see are things

  160. 6:46

    like the actual intent, the team

  161. 6:50

    conventions, past decisions, things that

  162. 6:53

    you've discussed in Slack, for example,

  163. 6:56

    uh architecture rationale, and so on.

  164. 6:59

    And that's why your agents need a

  165. 7:02

    context engine to get real work done.

  166. 7:05

    So, I'm going to now uh attempt a live

  167. 7:08

    demo. And hopefully the demo gods are

  168. 7:10

    kind. Um, so I want to pop back up

  169. 7:14

    conceptually. Oops, I think I'm on the

  170. 7:16

    wrong tab. We'll get to that one in a

  171. 7:17

    sec.

  172. 7:20

    So for now,

  173. 7:24

    sorry about that. And here we are.

  174. 7:29

    So I'm going to ask a question as if I'm

  175. 7:32

    a, you know, I'm a human and I want to

  176. 7:34

    get some information about my codebase.

  177. 7:38

    And, you know, the human layer hasn't

  178. 7:39

    gone away. We talk about agents and

  179. 7:41

    their need for context, but um humans

  180. 7:44

    are still asking questions about the

  181. 7:46

    codebase and we need that level of

  182. 7:47

    understanding because ultimately the

  183. 7:49

    accountability stops with us. When you

  184. 7:52

    hit merge on a PR, you need to

  185. 7:53

    understand what it's doing um and you

  186. 7:55

    need to understand how the architecture

  187. 7:57

    works. So this question I asked here um

  188. 7:59

    is about an internal component of our

  189. 8:02

    system called the source mark engine and

  190. 8:04

    you can see that it uh is able to

  191. 8:07

    articulate it fairly well. um

  192. 8:09

    understands the architecture. This this

  193. 8:11

    diagram here is uh is generated. So it

  194. 8:15

    this diagram doesn't exist. Um it just

  195. 8:19

    figures it out based on the um the way

  196. 8:22

    the code operates today and then some

  197. 8:24

    proposals for future architecture.

  198. 8:27

    And then uh what's really important is

  199. 8:29

    that you show your work. This is a trust

  200. 8:31

    building thing more than anything, but

  201. 8:33

    it allows people to see if um if the

  202. 8:37

    answer is maybe not entirely correct,

  203. 8:39

    then you can in look into the uh the

  204. 8:42

    knowledge base that you have and make

  205. 8:44

    corrections.

  206. 8:45

    Increasingly agents are doing this for

  207. 8:47

    you.

  208. 8:49

    So now um what I want to show you is

  209. 8:52

    another place where humans spend their

  210. 8:54

    time which is in Slack and this is where

  211. 8:58

    a lot of the decisions get made of

  212. 8:59

    course.

  213. 9:00

    So I can do something like this.

  214. 9:04

    And uh unblocked will sit and kind of

  215. 9:06

    listen for things that are things that

  216. 9:08

    can chime in on when it provides a high

  217. 9:10

    degree of Oh, sorry. We went to the

  218. 9:12

    wrong You guys can't see that. Thank

  219. 9:15

    you, Claire.

  220. 9:18

    Oh, come on down. Let's see if I can

  221. 9:20

    bring it up. There we go.

  222. 9:24

    Perfect. So I can ask questions like

  223. 9:27

    this in unblocked and if it thinks it

  224. 9:29

    can chime in on the answer then it will

  225. 9:31

    chime in. Otherwise I can just

  226. 9:34

    um address unblocked directly and ask

  227. 9:37

    the same question

  228. 9:41

    and when it thinks that it has an answer

  229. 9:43

    to give then it will give an answer and

  230. 9:46

    so we can get um

  231. 9:49

    quite a bit of interesting content there

  232. 9:51

    from unblocked. Thank you. Unblocked.

  233. 9:55

    I'm going to switch up

  234. 9:58

    and show you the the really interesting

  235. 10:00

    thing which is the agents. Okay. So, um

  236. 10:03

    in in that question, the source mark

  237. 10:05

    engine, I'm not sure if people picked

  238. 10:06

    up, but there was a little thing at the

  239. 10:08

    bottom there that said, you know,

  240. 10:09

    there's some optimization opportunities.

  241. 10:11

    Um so what I did here is I went into

  242. 10:13

    claw code and I asked it um without

  243. 10:16

    using unblocked to um

  244. 10:19

    uh generate uh a plan to optimize the

  245. 10:23

    source mark calculator and it did that

  246. 10:25

    and it happily went and you know

  247. 10:27

    searched through the code and and tried

  248. 10:28

    to figure out how the algorithm works

  249. 10:30

    and so on. Um and it it reached a

  250. 10:33

    conclusion that's great you know it does

  251. 10:35

    a pretty good job um but you know it

  252. 10:38

    maybe could do a little bit better. So,

  253. 10:40

    I asked that question again uh using

  254. 10:42

    unblock this time and it it really kind

  255. 10:46

    of nails the the nuances because it

  256. 10:48

    picks up on the the uh PRs that we um

  257. 10:53

    where we discussed future possibilities

  258. 10:55

    for improvement. um some Slack

  259. 10:58

    conversations that we had and uh of

  260. 11:01

    course you know notion and architecture

  261. 11:03

    documents and it shows its work and this

  262. 11:05

    is really important because um all of

  263. 11:08

    these things here the sources come back

  264. 11:10

    to Claude and then Claude knows exactly

  265. 11:13

    where to jump to next if it needs to

  266. 11:15

    elaborate on that context. And so I just

  267. 11:18

    want to show you what the impact of that

  268. 11:19

    is. So if I um Whoops.

  269. 11:24

    Thank you. If I pull up usage here, you

  270. 11:26

    can see that with unblocked, uh, the

  271. 11:29

    total cost was, you know, subd dollar to

  272. 11:31

    create the plan. Uh, took about a

  273. 11:33

    minute. Ignore the wall clock time

  274. 11:35

    because I've had this open for about an

  275. 11:36

    hour. But, um, it's about a minute. And

  276. 11:40

    then if I look at um the usage without

  277. 11:43

    unblocked, you can see that it's about 2

  278. 11:46

    minutes. And and and it costs more to

  279. 11:49

    generate all that context. Now, the

  280. 11:50

    reason that happens is because it has to

  281. 11:52

    do more work. It has to look around. has

  282. 11:54

    to discover things. Um, and this

  283. 11:56

    compounds, not only does it have to do

  284. 11:59

    more work to discover things, it doesn't

  285. 12:01

    discover the right things. So, when you

  286. 12:03

    get further down in your execution, it

  287. 12:06

    may be operating on the wrong plan or

  288. 12:08

    the wrong assumptions. And then you have

  289. 12:09

    to go back and you have to loop over and

  290. 12:11

    over again. So, the real value of a

  291. 12:13

    context engine is not like the upfront

  292. 12:15

    cost on these short tasks. It's the

  293. 12:18

    compounding effect. Um the the other

  294. 12:20

    Tariq from Sonar mentioned this in the

  295. 12:23

    keynote this morning and it's true like

  296. 12:25

    the loops compound and you have to be

  297. 12:27

    like um uh efficient the entire way

  298. 12:30

    through with your context. I'm just

  299. 12:32

    going to jump back to

  300. 12:35

    Safari and I'm going to point out um

  301. 12:37

    some really interesting things. So we

  302. 12:41

    also have a a code review agent.

  303. 12:44

    And when we say um you know

  304. 12:47

    organizational context, we're talking

  305. 12:50

    about more than just the underlying

  306. 12:52

    data. Uh we're talking about real

  307. 12:54

    intelligence. So what unblock does is it

  308. 12:58

    looks at um not like it looks at pull

  309. 13:00

    request data and there are other data

  310. 13:02

    sources for this and it generates a

  311. 13:05

    series of best practices that help align

  312. 13:08

    agents to your codebase. But we thought

  313. 13:10

    that this would be really helpful to

  314. 13:11

    surface for the review agent as well. So

  315. 13:14

    what you can see here is um

  316. 13:18

    it unblock chimed in and then Richie

  317. 13:21

    here said, "Oh, that's cool. That's

  318. 13:22

    something I would say." And that's

  319. 13:23

    because that actually was something he

  320. 13:25

    said. So it surfaced the uh the previous

  321. 13:28

    comments. Richie's one of the senior

  322. 13:30

    engineers and we use the sort of

  323. 13:33

    seniority or expertise as a signal um to

  324. 13:37

    boost uh comments that are important.

  325. 13:40

    Okay.

  326. 13:42

    So another uh interesting interaction by

  327. 13:45

    Richie, he uh discovered that the number

  328. 13:48

    of code review issues that were being

  329. 13:50

    surfaced dropped uh precipitously

  330. 13:53

    and he was debugging it with unblocked.

  331. 13:56

    Um he got all the way to the bottom and

  332. 13:58

    realized what roughly what the problem

  333. 14:00

    was and then asked unblocked to fix it.

  334. 14:02

    Now this this is something that we have

  335. 14:04

    internally um you know that we're

  336. 14:07

    experimenting with. Um, so unblocked uh

  337. 14:10

    can run as an agent in the cloud. Um,

  338. 14:13

    but what's really cool about this is

  339. 14:16

    that it has all your organizational

  340. 14:17

    context at its fingertips and the

  341. 14:20

    results are are pretty magical. So it

  342. 14:23

    can do things like generate this PR um,

  343. 14:26

    and then what you'll see here is that

  344. 14:28

    not only does it generate the fix, it

  345. 14:31

    also is able to relate it to the all the

  346. 14:33

    conversations that were happening. So

  347. 14:35

    this PR was created because and you read

  348. 14:37

    that context thing. It's mind-blowing.

  349. 14:39

    After this PR, we switched to uh Claude

  350. 14:43

    48 and it dropped a ton in issues

  351. 14:46

    because of the behavior is quite a bit

  352. 14:48

    different. So then it said Richie

  353. 14:50

    directly correlated the drop. Now what's

  354. 14:52

    this thing here? Let's click on it. It

  355. 14:54

    is a Slack conversation. So, it found

  356. 14:56

    the Slack conversation, correlated all

  357. 14:59

    of that, you know, past history back

  358. 15:01

    again, and then we ended up with a with

  359. 15:03

    a final PR.

  360. 15:08

    So, um, I'm going to I've got only a few

  361. 15:10

    minutes left. I'm just going to close

  362. 15:12

    this out really quickly. We have a uh a

  363. 15:14

    couple of open- source projects that are

  364. 15:16

    kind of interesting if people want to

  365. 15:17

    play with them. One is the document

  366. 15:19

    query engine. That was, uh, something

  367. 15:21

    that I talked about on Monday in my

  368. 15:23

    workshop. Um, I may uh talk about it

  369. 15:26

    again tomorrow, but I just want to give

  370. 15:27

    folks a sense of what this thing does.

  371. 15:30

    Um, whoops.

  372. 15:32

    If you want to play with it, it's open

  373. 15:35

    source, so you can just download it and

  374. 15:36

    have it go. It basically runs over your

  375. 15:39

    um uh GitHub repository, ingests uh your

  376. 15:43

    your historical pull requests, and then

  377. 15:46

    uh synthesizes a schema based on the

  378. 15:48

    documents that it can sample. Um and

  379. 15:51

    then from there you can issue any kind

  380. 15:52

    of queries that you like and get all

  381. 15:54

    kinds of insights out of it through the

  382. 15:57

    agent chat. You can ask all kinds of

  383. 15:59

    questions. Um and then lastly the

  384. 16:02

    engineering social graph. So this is the

  385. 16:04

    thing that I was talking about earlier

  386. 16:06

    that helps us pin down expertise and

  387. 16:09

    team relationships. Um so what you can

  388. 16:12

    see here is this sort of like the rough

  389. 16:14

    breakdown of our team structure at

  390. 16:16

    Unblocked. As you can see we're a fairly

  391. 16:18

    small team. Um and so we've got these um

  392. 16:23

    uh these clusters of people and how they

  393. 16:25

    relate to each other indicates the kind

  394. 16:27

    of um review relationships that they

  395. 16:30

    have. So these are you know these lines

  396. 16:32

    show like we review each other's code.

  397. 16:35

    Um

  398. 16:36

    we can then cluster that and generate

  399. 16:39

    team labels for that or show the

  400. 16:42

    coverage across your codebase. This is

  401. 16:44

    really cool. you can see kind of where

  402. 16:45

    the holes are, where you might be

  403. 16:47

    lacking expert coverage. Um, and that's

  404. 16:50

    exactly what we use within the context

  405. 16:52

    engine itself.

  406. 16:54

    All right,

  407. 17:00

    one last thing we have uh for those that

  408. 17:03

    want a taste of what a context engine

  409. 17:05

    can do but don't want to sign up for

  410. 17:07

    unblocked right away. Um you can use uh

  411. 17:11

    something that we call the context

  412. 17:12

    engine simulator which will basically

  413. 17:15

    build up a context behind the scenes on

  414. 17:18

    a per task basis and then use that

  415. 17:21

    context uh to to drive the task. It'll

  416. 17:24

    do it with context and without context

  417. 17:26

    so that you can see what the differences

  418. 17:28

    might be.

  419. 17:30

    This is a QR code for that if you want

  420. 17:33

    to just take a quick snap.

  421. 17:38

    Awesome. And I'll just land

  422. 17:41

    on a quote from one of our customers.

  423. 17:45

    50% fewer tokens, faster triage, better

  424. 17:49

    answers. And that's exactly what a

  425. 17:51

    context engine can do.

  426. 17:54

    One last shout out um before we end. My

  427. 17:57

    colleague Brandon is giving a talk in

  428. 18:01

    10 minutes uh at room 2020. um he's

  429. 18:04

    going to speak to in a lot more detail

  430. 18:06

    about some of the higher level things

  431. 18:08

    that context engines can do. I'm going

  432. 18:10

    to run over there right after this and I

  433. 18:11

    think all of you should follow me.

  434. 18:14

    Awesome. Oh, and don't forget to get a

  435. 18:16

    coconut.

  436. 18:33

    >> [music]