AI Engineer World's Fair 2026
How to Get Your Org to Adopt Coding Agents (Without Shipping Garbage) — Eyal Blum, Figma
Read the talk
How to Get Your Org to Adopt Coding Agents (Without Shipping Garbage)
Eyal Blum explains Figma’s still-evolving approach: turn recurring failures into verification, give engineers substantial planning work to own, and treat skeptical reviewers and scarce human attention as design constraints.
From a talk by Eyal Blum
At a glance
Ideas worth remembering
Early coding-agent wins do not generalize automatically. Dependable adoption begins after larger tasks expose missing context, guardrails, and verification.
Turn recurring agent checks into deterministic tests, establish verification criteria before implementation, and reserve human review for functionality, intent, and product judgment.
A plan should preserve a stable purpose, divide work into reviewable phases, and validate each phase before later work depends on it.
Experienced skeptics often carry the codebase knowledge that agent workflows lack. Give them ownership of the safeguard roadmap instead of treating resistance as a persuasion problem.
Protect reader attention by placing a concise human-written explanation before clearly distinguished generated detail, and state what expert judgment you need from the recipient.
Adoption can begin inside familiar workflows such as Slack, but incomplete automation and build-system dependencies remain real constraints.
Adoption starts fast, fails hard, then becomes engineering
This is a talk about Figma’s internal engineering organization, not its product. The problem is how teams can adopt coding agents without lowering the quality of an established codebase. Blum’s starting model has three acts: simple tasks produce an apparent “10x faster” moment; the same techniques fail on larger work, creating bugs and breaking trust; then users learn the harder discipline of supplying context, guardrails, verification, and better instructions.
Teams do not move through those acts together. Some at Figma have already changed their workflows around agents, while others are still experimenting or recovering from bad results. They nevertheless contribute to the same product. Adoption therefore cannot depend on an organization-wide leap of faith; teams at different levels of confidence must coexist while the organization improves the conditions under which agents can work reliably.
Initial experiments can appear dramatically faster.
Early speed creates confidence, larger tasks expose failure modes, and dependable use emerges only after teams engineer the surrounding workflow.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The costs arrive before the safeguards
The first cost is reduced developer agency. Engineers who enjoyed writing code and entering a state of flow can end up in a prompt-and-wait cycle: ask the agent, wait for output, inspect it, and ask again. Blum reports lower job satisfaction and possible burnout. Faster implementation does not automatically make the work more rewarding for the person directing it.
The second cost lands on the engineers who understand the codebase best. They know its undocumented assumptions and recognize where generated changes can fail. That institutional context acts as “mental duct tape”: they stop bad changes, absorb more review work, become bottlenecks, and grow frustrated. In Blum’s account, the strongest engineers may adopt agents last because they encounter the risks first.
The third cost is attention. Blum describes design documents, Slack messages, and emails becoming three or four times longer, while email volume grows two or three times, without a comparable increase in substance. Readers must work harder to identify what matters and what deserves trust. These are observations from an unfinished transition: Figma has seen progress, but Blum explicitly says the organization has not reached the other side.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Move verification down the stack before generating more code
Blum calls verification the highest-value investment a team can make in its codebase. The immediate move is to shift checks away from manual exploration and toward tools an agent can use itself. He cites Playwright with MCP as an example: giving an agent a way to explore an application reduced work that previously required a person and unlocked productivity for several teams. The important mechanism is closed-loop work—the agent can inspect the environment and gather evidence before asking for review.
Agent reasoning should not remain the permanent implementation of a check. When an agent discovers a useful, repeatable test, encode it as deterministic automation. That saves time and tokens, makes results reproducible, and reserves the language model for questions that still require judgment. The progression is: use the agent to discover what needs checking, then turn known checks into ordinary software.
Test order matters too. Blum recommends a red-to-green TDD cycle: establish a failing test and then ask the agent to implement code that passes it. If the agent writes the implementation first and the test afterward, it can fit the test to what it already produced. Test-first work instead fixes the verification target before implementation begins. Blum reports that this usually improves results, though it does not guarantee that either the test or the implementation is correct.
The resulting review pyramid has three layers:
- Deterministic checks: Linters, compiler checks, and unit tests cover conditions software can decide directly.
- Agent review: Explicit criteria can cover encoded architectural standards and other checks that need some reasoning.
- Human judgment: People concentrate on functionality, intent, and whether the change is the right thing to build.
The goal is not to remove human review. It is to stop spending scarce human judgment on checks that machines can perform consistently.
Assess functionality, intent, and whether the change should exist.
Push routine checks toward deterministic analysis and preserve human attention for product and functional judgment.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Plans give engineers decisions to own
Planning addresses both reliability and the loss of developer agency. Rather than spending the day reacting to generated code, engineers can own the decisions: what to build, why it matters, how the work should be divided, and how each part will be checked. Blum says it is not unusual to spend a week writing and revising a detailed plan, resolving choices, and getting teammate review before an agent begins implementation. That preserves deliberate engineering work while moving mechanical execution to the agent.
A useful plan begins with a stable explanation of why the work exists. Blum recommends a prominent executive summary that the agent can revisit when it starts to drift—and that the agent should not casually rewrite. The plan then breaks the work into small phases that can each be verified independently.
His sizing heuristic is deliberately human: would he want to review the corresponding pull request in one sitting? If a reviewer feels they need to fetch coffee before opening it, the phase is too large. This matters because an agent’s ability to generate a large diff does not increase a reviewer’s ability to understand it.
Every phase also needs a validation gate or explicit exit criteria. Otherwise, a five-stage plan can complete its first stage without checking it and build four more stages on a faulty assumption. Independent validation prevents an early mistake from becoming the foundation of later work. Once this structure exists, teams can choose their preferred agent loop; Blum sees diminishing returns in forcing everyone onto one centralized workflow, provided collaborators can still inspect and iterate on the work.
Blum illustrates the approach with roughly 20 pull requests, some around 10 lines and others around 100, with probably nothing larger. He describes about a week of planning, another week aligning with three teams, and an overnight agent implementation. He then estimates that roughly six weeks of coding work took one week and calls the result a 5x speedup after accounting for review. He also says the example probably came from two plans, so the spoken timeline should be treated as an approximate project account rather than a controlled productivity measurement. The durable lesson is simpler: implementation speed only counts if the resulting changes remain small enough to review, and review time belongs in the calculation.
An executive summary explains why the work exists and anchors later decisions.
Each phase must pass its own validation gate before later work can safely depend on it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Treat skeptics as the safety roadmap
Blum returns to the experienced engineers carrying the largest review burden. Their objections identify missing validation, undocumented assumptions, and places where agent tooling fails. Trying to persuade them to use AI misses the information inside their resistance. Their complaints are a roadmap for making agent interaction with the codebase safer.
His organizational recommendation is to give those engineers ownership of that roadmap. As they encode safeguards and remove recurring failures, their own review burden should fall; adoption can follow from seeing the system improve rather than from an executive mandate. Blum notes that less than an hour of brainstorming with a group produced a substantial list of fixes, suggesting that the missing-safeguard backlog may already be sitting in reviewers’ heads.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Generated communication needs an attention budget
Once generation makes text cheap, reader attention becomes the constrained resource. Blum’s team uses a simple convention to help allocate it: distinguish what a human wrote from what AI generated. The distinction tells readers where the author has exercised direct judgment and where they should expect more uncertainty or noise. Blum presents this as a team practice being considered more broadly, not as an organization-wide Figma rule.
Every pull-request description on his team begins with a short human-written explanation of what the change does. Generated detail follows below. The author may read and edit that material, but did not personally write every line, so readers should give the opening explanation greater weight and scrutinize the generated section more carefully. The same convention can apply to Slack and email: disclose the generated portion and tell people where to focus.
The practice came from a failed interaction. Blum used AI to analyze comments from skeptical senior engineers, then sent the analysis without clearly separating his contribution from the generated text. They regarded the result as unexpectedly sloppy. His correction was not merely to add an AI label; he also needed to state his intent. He lacked enough context to judge the analysis and wanted their expert feedback. That framing would have told recipients both what was uncertain and what work he was asking them to do. For Blum, changing communication culture is as important as solving the engineering problems.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Let people meet agents inside familiar work
Blum’s final tactic is intentionally less sophisticated than a software factory: put agents where people already work. In a Slack conversation, someone can tag an agent, ask it to perform a task, and have it return the result in the same thread. That removes the setup cost of learning a separate interface and turns adoption into a shared, concrete experience.
The social delivery matters. With a colleague who is not convinced, Blum suggests trying the agent together without making the gesture passive-aggressive. If the agent closes the loop successfully, that result can encourage the colleague to try it independently elsewhere. More elaborate automation can come later; the first useful experience should happen inside an existing workflow.
The talk ends with limits still visible. Figma continues to experiment, its automation is incomplete, and dependencies in parts of its build system still complicate effective use of cloud agents. Shipping external AI features has not made internal adoption a solved problem. Blum describes the transition as both cultural and technical—and, after 15 years working in the Valley, as the largest change he has seen by orders of magnitude. His recommendations are therefore working practices from an organization still learning, not a finished operating model.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The official AI Engineer page with the recording, timestamped transcript, chapters, and reading version.
Related talks
- Making Codebases "Agent-Ready"
Extends Blum’s verification argument with concrete codebase-readiness measures such as opinionated linters, tests, documented interfaces, and continuous validation.
- No Vibes Allowed: Solving Hard Problems in Complex Codebases
Develops the research-plan-implement workflow for established codebases where undocumented context and rework make agent use difficult.
- Your Attention Is the Bottleneck, Not Your Agents — Zack Proser, WorkOS
Explores the same scarce-attention constraint through layered verification, background agent work, and developer well-being.
Read the complete timestamped transcript
- 0:01
[music]
- 0:12
>> Good afternoon. My name is Alon Blum. I
- 0:15
am a software engineer at Figma.
- 0:17
And in my talk today, we're going to
- 0:19
talk about how we've adopted or are
- 0:22
adopting
- 0:24
agent into our workflow at Figma
- 0:27
while maintaining
- 0:28
high quality for our code base.
- 0:31
So, as you may know, Figma is the
- 0:35
browser-based
- 0:36
editor where design and engineering and
- 0:39
now AI agent collaborate together to
- 0:42
ship code.
- 0:43
Uh this
- 0:45
Figma has
- 0:47
pivoted very strongly from being a
- 0:48
traditional tool to an AI-first tool.
- 0:52
But in this talk, I'm not going to talk
- 0:53
about our product. I'm going to talk
- 0:54
more about our internal organization and
- 0:57
how our engineering org has been
- 1:00
adopting AI agents.
- 1:04
Um what we we found internally is both
- 1:08
organizations, companies, and individual
- 1:11
there's kind of a three-act
- 1:13
process of AI adoption.
- 1:16
You start with picking up something,
- 1:18
whether it was a lot of the people in
- 1:20
this room who have been doing using our
- 1:23
AI pal and have been using AI for a
- 1:25
while and they picked up something and
- 1:27
got some simple things to work very
- 1:29
well.
- 1:30
10x faster.
- 1:32
Then you start applying those same
- 1:34
practices to bigger problems and AI
- 1:37
fails pretty badly at that, gives you
- 1:39
bad stuff,
- 1:40
lots of bugs,
- 1:42
and the trust that you build breaks
- 1:44
down.
- 1:45
And then
- 1:46
from that point, you start building the
- 1:48
real skill, which is learning how to use
- 1:51
AI correctly and put the right
- 1:53
guardrails and the right prompting and
- 1:55
the right context and all the stuff that
- 1:57
we've been talking all day about here in
- 1:59
all the talks in order to actually build
- 2:01
a real scale.
- 2:03
And one thing that
- 2:06
is happening internally as we we adopted
- 2:09
whether teams or individuals
- 2:11
the adoption is uneven. We have teams
- 2:14
that are very AI forward and have
- 2:17
already transformed their entire
- 2:18
workflows and then we have teams that
- 2:20
are still experimenting in the earlier
- 2:24
act and or have lost confidence and they
- 2:26
all need to work together in order to
- 2:27
ship our product. Um
- 2:30
So, they need to coexist
- 2:33
in the organization and we need to find
- 2:34
a way to support them and while bringing
- 2:36
on everybody along for the journey and
- 2:39
getting everybody to the third act of
- 2:41
the story.
- 2:44
Aside from that main friction point, we
- 2:46
have also noticed other friction points
- 2:48
that happened
- 2:50
as we adopt AI.
- 2:53
One thing that we've heard a lot from
- 2:55
developers and managers have have been
- 2:57
noticing is that reduced developer
- 2:58
agency causes
- 3:00
um
- 3:01
engineers to lose some of their job
- 3:03
satisfaction. So, if a lot of people
- 3:05
used to take a lot of pride and
- 3:07
enjoyment in writing code and getting
- 3:09
into the flow
- 3:10
and a lot of people feel like that's
- 3:12
been lost or they're losing a lot of
- 3:13
that
- 3:14
element and getting into more of a
- 3:16
prompt cycle where they just wait on
- 3:18
output from AI and then speak to the AI
- 3:20
that like not as much fun as they used
- 3:22
to have and they're getting burned out.
- 3:25
Um we've noticed another interesting
- 3:27
thing. It's actually our best engineer,
- 3:29
the one that hold all their contacts in
- 3:30
their brain. Um they end up getting out
- 3:33
of the burden. What ends up happening is
- 3:35
they they know where all the pitfalls
- 3:37
are. They are like holding together with
- 3:40
with like their mental duct tape all the
- 3:42
places
- 3:43
that agents are not working well and
- 3:45
they're preventing all the really bad
- 3:47
stuff from coming in or all they they
- 3:49
have all the institutional contact that
- 3:52
have never written down in their head
- 3:54
and they get so much burden and and
- 3:56
become bottlenecks and gets really
- 3:58
frustrated. So, they actually end up
- 4:00
being slowest to adopt because they see
- 4:01
all the problem
- 4:03
uh first hand.
- 4:04
That's another big big issue that we've
- 4:06
seen.
- 4:08
Um and this one I'm sure everybody can
- 4:10
resonate or in Sorry, I'm sure everybody
- 4:14
here will resonate.
- 4:15
Um that all of a sudden all the design
- 4:17
docs and all the Slack messages you
- 4:20
know, this the emails have gotten three
- 4:22
or four times as long and we've gotten
- 4:23
two or three times as many emails
- 4:26
and they say basically as much as they
- 4:28
did before. So, communication has gotten
- 4:31
quite inefficient and some of the
- 4:33
markers of like what is high quality and
- 4:35
important things versus not so much high
- 4:37
quality
- 4:38
um has become challenging to navigate.
- 4:42
Um so, I'm going to spend uh the next
- 4:45
few minutes talking about some of the
- 4:47
lessons that we've learned and how we've
- 4:49
been trying to apply this. This is a
- 4:51
journey. We have not come out through
- 4:52
the other end, but we've seen some
- 4:54
really interesting progress
- 4:55
along a lot of these lines.
- 5:00
Um I think this
- 5:02
uh a lot of the speakers here have
- 5:04
touched upon this, but investing in
- 5:06
verification is probably the highest
- 5:08
value thing we can do in our code base.
- 5:10
Um anytime that we can lift a
- 5:14
left shift anything in our workflow from
- 5:16
a human needing to do it to an agent
- 5:18
being able to verify it.
- 5:20
So, for example, when uh Playwright and
- 5:24
MCP came out, instead of having humans
- 5:27
navigate the code, now the agent can
- 5:28
explore the code. That was a big win
- 5:31
unlock for productivity in a lot of our
- 5:33
team. That's really That's always a
- 5:36
a big win for us.
- 5:38
The other thing is
- 5:40
um
- 5:41
it's even better if when you find
- 5:43
something that the agent has found to be
- 5:45
useful,
- 5:46
take the time to take that and encode
- 5:49
into a deterministic flow.
- 5:51
A deterministic flow that can be easily
- 5:53
repeated is saved on tokens, save on
- 5:55
time for the and then it also you also
- 5:57
know that you're using the the LLM when
- 6:00
it needs to reason, but when you have
- 6:03
something that is already
- 6:05
known and basically can be encoded into
- 6:07
a test, spending that time always always
- 6:09
pays dividends.
- 6:11
Um
- 6:12
And another tip, if you tell your scale
- 6:15
your agent to write the code that you're
- 6:17
writing
- 6:18
um like at the red to green to red to
- 6:21
green at the TDD style,
- 6:24
it almost always gives you better
- 6:25
results because you set a goal, then you
- 6:28
tell the agent to strive toward that
- 6:29
goal, it will almost always give you
- 6:31
better results than writing the code and
- 6:33
then writing the test afterward because
- 6:34
then it will fit the test to the code
- 6:37
rather than fit the code to pass the
- 6:38
verification criteria.
- 6:42
Um this is the testing pyramid that uh
- 6:44
can't the classic testing pyramid from
- 6:46
the
- 6:47
previous
- 6:48
uh just when you think about the testing
- 6:50
themselves, which you had the end-to-end
- 6:52
test and the integration test and the
- 6:54
unit test. This is very similar.
- 6:57
Move as much as you can down to the
- 6:59
deterministic analysis where that's
- 7:00
linting, the compiler, um the unit test
- 7:04
themselves.
- 7:05
Whatever that can come be covered
- 7:07
easily, you can have engine agent do
- 7:09
reviews on it based on on criteria, so
- 7:12
um
- 7:13
architectural standards that that have
- 7:15
been easily encoded into the code base,
- 7:18
you can move into the agent. And then
- 7:20
only at the very top you need to have
- 7:22
some sort of human review, which is
- 7:24
usually around the functionality and
- 7:26
this is the right thing to build.
- 7:28
That like only leave the human to do
- 7:30
what the humans need to actually be
- 7:31
involved in.
- 7:34
Um Another really important thing is the
- 7:37
planning
- 7:38
versus prompting. This is really tied
- 7:40
into the giving agency back to
- 7:42
developers and finding a replacement to
- 7:45
the craft of writing code.
- 7:48
Um
- 7:50
spending a lot of time writing the plan
- 7:52
and then
- 7:53
sending enough to the agent basically as
- 7:54
a
- 7:55
as an implementation that can be done
- 7:57
automatically is something that we find
- 8:00
to really
- 8:02
kind of
- 8:03
reintroduce the joy of of building back
- 8:06
into the process.
- 8:07
And so it's not uncommon to spend a week
- 8:11
writing a very detailed plan, making all
- 8:13
the decisions, flushing it out,
- 8:14
iterating, sending it out to teammates
- 8:16
to review.
- 8:18
And then only when it's ready and you've
- 8:19
flushed out all the decision, you can
- 8:21
send it to the agent. The agent will
- 8:23
um send it back to you when it's
- 8:25
implemented.
- 8:26
And that that has been really successful
- 8:29
also in accelerating and also
- 8:32
really restoring some of the joy into
- 8:35
the development process.
- 8:38
Uh so what makes a good plan? Um
- 8:41
really important to start with a why at
- 8:43
the top. It really helps preventing
- 8:45
agent drift. If you have like a bold big
- 8:47
section of kind of like it when you
- 8:48
write a design doc, you want to have the
- 8:50
executive summary. Put that in there for
- 8:52
the agent. Otherwise, they'll start
- 8:53
drifting over time and make sure that
- 8:55
the agent don't go back and change that
- 8:57
because they feel like it.
- 8:59
Uh so we start with a why.
- 9:01
Make sure that the plan can be broken
- 9:02
down
- 9:04
into small parts that can each be
- 9:06
verified independently.
- 9:08
And my personal way of knowing what is a
- 9:12
good size would I want to review that
- 9:14
the PR that will correspond to that
- 9:15
part? If it's going to be too big for me
- 9:17
to want to review in one sitting, it's
- 9:18
kind of like the test is
- 9:20
I'm going to get need to get a cup of
- 9:21
coffee before I read this.
- 9:23
That means it's too big and I'm going to
- 9:24
want to have it broken down into pieces.
- 9:27
And then
- 9:29
I make sure that each part can be
- 9:30
validated independently cuz what I don't
- 9:32
want to have is
- 9:34
have five stages and then the first one
- 9:37
is written but not validated, and then
- 9:38
everything else is is built on top of
- 9:40
all the assumptions. So, having kind of
- 9:43
a validation gate or an exception
- 9:44
criteria for each phase really helps and
- 9:48
make the plan uh resilient to drift. And
- 9:52
all and and there's all kind of
- 9:53
technique on how to manage the contacts
- 9:55
and
- 9:56
doing a a software factory on top of
- 9:59
that. But once you have the plan, you
- 10:00
can use whatever loop uh you want or
- 10:03
whatever workflow you want
- 10:05
in order to implement it.
- 10:07
Uh
- 10:08
this is a screenshot that I randomly
- 10:10
picked up a plan, but this is what I
- 10:11
usually look for. The executive summary
- 10:13
at the top, the phases break it down,
- 10:15
and then each one of them I would go
- 10:17
into lots of details so that I can just
- 10:18
fit it into a sub agent, and the sub
- 10:21
agent can independently work on that and
- 10:23
not have to worry about it. Um that's
- 10:26
that's it. There are other workflows
- 10:27
that would work or other structure to
- 10:29
the plan. I find that
- 10:32
part of the things that great about uh
- 10:34
AI workflows is that everybody can set
- 10:36
up the thing that works best for them.
- 10:38
Oh-oh.
- 10:41
No, thank you.
- 10:43
Everybody can very easily set up the
- 10:45
workflow that work exactly for them for
- 10:47
them. So, there's
- 10:49
diminishing return in trying to
- 10:50
centralize everybody on one thing, but
- 10:51
as long as it works for their flow and
- 10:53
other people can iterate with them, I
- 10:55
find that it generally works very well.
- 10:59
And this is just an example kind of a
- 11:01
brag of like this is uh could be a
- 11:03
result from a plan. Um there are
- 11:06
probably 20 PRs here. Some of them would
- 11:08
be maybe 10 lines, and some of them
- 11:10
would be 100 lines. There's probably
- 11:11
nothing bigger than that, and that
- 11:13
allows us to
- 11:15
This is in the pre-AI world, this plan
- 11:18
probably worked in that for a week. I
- 11:20
aligned with the other with three other
- 11:21
teams for another week on that, and then
- 11:23
I just sent it to an agent to implement
- 11:25
overnight, and it came back. This is
- 11:27
probably from two plans, not one, but
- 11:29
it's it's basically six weeks of of
- 11:32
coding work just It's
- 11:34
um
- 11:35
only took 1 week, so that's where
- 11:37
I got the 5x speed up. If I include the
- 11:41
review cycle at the end that we always
- 11:42
have to remember.
- 11:45
Um moving on from planning
- 11:49
back to the issue that we had with the
- 11:51
skeptics and the people who are burdened
- 11:53
with the most work, make sure that
- 11:55
you
- 11:56
bring them in and take their feedback
- 11:58
really seriously.
- 11:59
They're skeptic because they're seeing
- 12:01
the the way you are lacking validation,
- 12:03
where your tools fail. So, and their
- 12:07
feedback is basically the road map of
- 12:08
how to improve your agent
- 12:10
interacting with the code base.
- 12:12
So, just make sure to bring them in
- 12:15
rather than trying to
- 12:16
um figure out how to make them use the
- 12:18
AI. Just let's have them be in charge of
- 12:21
the road map to
- 12:23
make AI safe your organization, and they
- 12:25
will come along once they see that
- 12:28
that the improvement that they're making
- 12:30
actually making their life better.
- 12:33
Um and as you can see, they'll not be
- 12:35
shy about telling you what you need to
- 12:36
fix. This is Latin hour sitting with a
- 12:39
bunch of people, and
- 12:41
this is the result of brainstorms.
- 12:45
Um another thing that's
- 12:47
been really helpful with my team
- 12:49
specifically, and we're working to adopt
- 12:51
it
- 12:52
in the broader organization as well, is
- 12:54
to make sure that you have an
- 12:56
attention-aware communication.
- 12:58
In the age of AI, human attention is a
- 13:00
scarce resource. I think I've heard it
- 13:02
for multiple talks, and a lot of people
- 13:03
have
- 13:04
have come to the same conclusion. You
- 13:06
can't get more human attention. So,
- 13:08
where you spend your time and what
- 13:09
you're reading is really becomes really
- 13:11
important.
- 13:13
Um so, since it's such a scarce
- 13:16
resource,
- 13:17
marking what was generated by AI versus
- 13:19
what was written by human is really
- 13:21
helpful to know how much time you need
- 13:23
to spend reading this, and how much slop
- 13:26
can you expect in this part of the
- 13:27
communication?
- 13:29
Um
- 13:31
and that can building a
- 13:33
new culture around that
- 13:34
self-communication that really helps. Um
- 13:37
So, for example,
- 13:39
um the team that team that I work with,
- 13:41
we've decided we always every PR
- 13:44
description will start with something
- 13:45
like that, something that I wrote by
- 13:47
hand. It could be very short that I
- 13:48
describe what this is in code and what
- 13:50
this is doing. And then the AI
- 13:52
description is going to come after that,
- 13:54
which is I will probably read it. I will
- 13:56
probably edit it to remove uh some wrong
- 13:58
things, but they didn't write every line
- 14:00
here, so they should be more suspicious
- 14:02
and they should pay more attention to
- 14:04
what I wrote in the top and they should
- 14:05
override it. Things like that in Slack,
- 14:08
in email, it's like leaning into the
- 14:11
fact that everybody knows that you're
- 14:12
using AI to to craft your communication,
- 14:16
but just let them share about it, tell
- 14:17
them what they should read and what you
- 14:20
they should pay less attention to.
- 14:22
And I remember early on, maybe
- 14:25
like earlier in this year, I tried to I
- 14:30
had some senior engineers in our org
- 14:32
that had kind of were very much AI
- 14:35
skeptic and I tried to reach out to them
- 14:37
to see what was the problem, what was
- 14:38
going on. I said, "I tried to run an
- 14:40
analysis on some of the PR comments that
- 14:42
you've run." And obviously
- 14:44
I used the AI to do that.
- 14:46
And then I didn't distinguish very
- 14:48
clearly what I wrote versus what they
- 14:51
what AI generated. And they got very
- 14:54
upset. They're like, "Why is sending I
- 14:56
did not expect somebody um that I
- 14:58
respect this much to send me
- 15:01
something that's clearly this sloppy."
- 15:02
And then like I I took immediately like
- 15:05
I apologize. I realize I should have
- 15:07
marked it clearly and marked my
- 15:08
intention like this is what I wrote.
- 15:10
This is what the AI wrote and I need
- 15:11
your feedback on that because I don't
- 15:13
have the context to know if it is sloppy
- 15:14
or not and that's what I'm asking you
- 15:16
for, so lesson like that and change the
- 15:18
culture is just as important as some the
- 15:21
engineering challenges that we've been
- 15:24
facing.
- 15:28
Um another thing that's really helpful
- 15:30
around the adoption is
- 15:34
um as you progress through adoption,
- 15:36
there's a lot of very fancy tools and a
- 15:37
lot of very fancy workflow that we've
- 15:40
we've been implementing, but one of the
- 15:42
really effective thing is just letting
- 15:45
people use the AI where they're at. So,
- 15:48
uh it help it really helps normalize
- 15:51
uh the use of AI for everyday tasks and
- 15:53
it helps reduce the friction.
- 15:55
And really one of the most powerful
- 15:56
thing is being able to tag an agent in
- 15:59
the Slack message with somebody and they
- 16:01
can you just do this for me?
- 16:03
And have the agents to close the loop in
- 16:05
the thread.
- 16:06
Um that that's kind of thing is really
- 16:09
powerful. And then you can go on top of
- 16:11
that and have all this thing automated
- 16:13
and do all kind of fancy things, but if
- 16:15
you have a new conversation with
- 16:16
somebody who's not fully bought in and
- 16:18
then you can tag it in a non like
- 16:21
non-passive-aggressive way. You can tag
- 16:23
it and say, "Let's try to see if the
- 16:25
agent can get it this time."
- 16:27
And they close the loop and if it's a
- 16:28
good experience, that really helps
- 16:30
people try it out on their own
- 16:32
in other cases.
- 16:36
And our journey continues. We're still
- 16:38
learning even though we're shipping AI
- 16:40
externally, our AI adoption
- 16:43
um
- 16:44
we're experimenting with with so many
- 16:46
things all the time. Our automation
- 16:48
story is not
- 16:50
uh fully there yet. We're still trying
- 16:52
trying to figure out when we should use
- 16:54
how we can use cloud agent effectively
- 16:56
given all the dependencies we have for
- 16:58
some of our bell system.
- 17:00
And so we are continuing to learn. It's
- 17:02
a culture shift, it's an engineering
- 17:03
shift and I don't know about you, but
- 17:06
for I've been I've been working in the
- 17:08
valley for the last 15 years and this is
- 17:10
the biggest change by orders of
- 17:12
magnitude of everything that I've seen
- 17:14
in term culture and technology.
- 17:17
So, um we're all here together and we're
- 17:19
all figuring it out and that's that's
- 17:21
what I wanted to talk to you today.
- 17:23
Thank you.
- 17:24
>> [applause]
- 17:40
[music]