AI Engineer World's Fair 2026
AI-Native Organisations Run on Skills: How to Structure and Scale Them — Imad Touil, QuantumBlack
Read the talk
AI-Native Organisations Run on Skills: How to Structure and Scale Them
Imad Touil explains how reusable agent procedures become organizational infrastructure: specialized skills, shared catalogs, model-aware testing, and human owners who keep the whole system useful as it grows.
From a talk by Imad Touil
At a glance
Ideas worth remembering
The coding-agent sequence of specify, plan, task, and implement builds a product increment inside a larger lifecycle of strategy, discovery, data preparation, operations, and improvement.
Specialized, discoverable skills let teams reuse procedures and compose them into workflows. Progressive disclosure reduces context cost by loading relevant knowledge when needed.
A shared registry needs metadata, dependencies, versions, access control, evaluation, and observability—and named people responsible for maintaining its domains.
Skills can contain executable scripts and sensitive business logic. Security checks and retrieval permissions address different risks.
Retest skills against their tasks and changing models. Automatic evolution increases the need for governance because it accelerates changes to shared procedures.
Creating a skill is easier than maintaining it together
The hands drop as the questions get harder. Who has created and used skills? Who shares them with a team? Who governs and maintains them across an organization? Imad Touil, a distinguished engineer at QuantumBlack, sees only a few hands left for the last question. That gap sets up the practical problem: a useful procedure on one engineer’s machine still needs a way to become shared, maintained organizational knowledge.
The agentic software stack has two loops in Touil’s model. The inner loop is the coding agent and its harness—the runtime that manages context, tools and MCP connections, memory and state, and skill loading. The outer loop is the workflow that arranges skills, sub-agents, MCP servers, and hooks around the work. These layers serve different purposes: the harness supplies execution machinery, while the workflow shapes how that machinery behaves for a particular job.
Supporting infrastructure makes those loops usable across an organization. A sandbox supplies the execution environment. An MCP gateway manages tool access, and a model gateway manages and optimizes access to local or frontier models. A knowledge graph represents core IT systems, code, and shared registries. The context layer then brings the task’s ingredients together: project instructions, tool schemas that explain what can be called, conversation history, and retrieved files or code. Skills operate inside this larger system; they do not replace its tools, runtime, or context.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Specify, plan, task, implement builds one product increment
The familiar coding-agent sequence is straightforward: specify what to build, plan the design, break the work into tasks, and implement them. At organizational scale, that sequence builds a product increment. It sits inside a much longer journey from deciding what the business needs to operating something customers can use.
Product strategy comes first: define success metrics and a roadmap, using market research, competitive analysis, and customer interviews as inputs. Discovery turns that direction into problem statements, possible solutions, validation, experiments, and user stories. Those steps determine whether the proposed implementation is worth doing before a coding agent starts producing it.
Even then, development may depend on data work. A team might need to clean its data catalog, adjust endpoints or integrations with core systems, build a data pipeline, and validate data quality. Preparing those assets makes the product increment possible. After implementation, platform engineering supplies infrastructure and infrastructure-as-code modules; launch starts another phase of performance optimization and incident resolution, feeding the next cycle.
There is also more than one software development lifecycle within a company. Mobile applications, internal employee platforms, customer-facing products, and different departments can require different workflows. A universal workflow would have to absorb all those differences. Touil instead treats the organization as a collection of lifecycles, each with procedures that can be made explicit and shared where appropriate.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Skills carry the procedure that tools alone do not supply
Within the workflow, Touil assigns four components distinct jobs:
- Hooks: Trigger actions at events along the workflow.
- MCP servers: Expose tools, often supplied by products the organization already uses.
- Sub-agents: Execute delegated tasks in separate contexts, reducing pressure on the main context window.
- Skills: Carry the organization’s know-how about how a task should be performed.
This division explains why connecting tools and delegating work can still leave a workflow underspecified. The agent needs the procedure as well as the means to act.
Workflows are “harness blueprints”: they shape the coding harness’s behavior at runtime. Skills make those blueprints more explicit by supplying repeatable task instructions and, where needed, scripts. Touil describes the intended result as deterministic. The supported mechanism is narrower: reusable instructions constrain the work, and a saved script can perform a specific operation consistently; neither establishes that every model decision or complete workflow result will be identical.
The adoption story moves from Anthropic’s introduction of skills to an open standard and uptake across agent harnesses. Touil’s snapshot of public repositories and registries illustrates growing creation, while his benchmark discussion reports better software-engineering and cybersecurity task outcomes with skills than without them. The comparison supplies motivation rather than an effect size: no scores or experimental conditions are given here, so it cannot tell a team how much improvement to expect on its own tasks.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Borrow the design discipline of microservices
The engineering problem looks familiar from microservices: many small capabilities need to remain understandable and useful together. A skill should describe one specialized task rather than grow into a monolith of everything an agent might need. Smaller responsibilities make it easier to reuse the procedure and combine it with other procedures without duplicating or conflicting instructions.
Several design properties work together:
- Reuse and modularity: Package a task so another team can use it without recreating the same procedure.
- Discovery: Make the skill findable by the team or agent that needs it.
- Portability: Use the shared standard to carry skills across workflows and supporting harnesses; Touil uses moving from Claude Code to Cursor as his example.
- Composition: Design responsibilities so multiple skills can run together without overlapping rules that fight each other.
- Consistency: Preserve a repeatable method for performing the task instead of asking every engineer to steer the agent from scratch.
Cost efficiency comes from progressive disclosure: bring in the relevant skill, in the amount needed, at the point when the task needs it. Loading an organization’s whole procedural library into every context would consume tokens regardless of relevance. Selective loading makes the library available without making every task carry all of it. Touil’s aim is to make organizational know-how executable and portable while reducing the context cost of using it.
A data-retention policy makes this concrete. When an agent manipulates customer data, it needs instructions about the applicable retention rules. That procedure is specialized enough to become one skill. It can then be reused wherever customer data is handled, rather than being rediscovered through a new conversation for each application.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Compose retention rules into a regulatory review
The retention-policy example expands into a regulatory disclosure review. Retention is only one responsibility. The catalog also contains disclosure standards, GDPR rules, and fill-in templates. A workflow pulls these skills at runtime so that a feature handling data can be reviewed against the relevant procedures across web, mobile, and other applications.
The visible change is in what the workflow produces: a storable audit report and identified improvements that can feed back into the codebase. The causal sequence is cataloged rules, runtime composition, review, then actionable output. The report gives the review a persistent result; the improvement findings connect it to development rather than leaving it as a separate compliance exercise.
How do separate policy skills become one review with a feedback path? The diagram shows the shared workflow as the meeting point for four specialized inputs. Its two outputs have different destinations: preserve the report, and use the findings to improve the code. Composition lets the rules stay separately maintained while contributing to the same task.
Instructions for handling customer data under retention rules.
Four specialized skills feed the regulatory disclosure review; the review produces a report and findings that return to development.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Ungoverned skills accumulate familiar technical debt
Once teams build skills independently, reuse can fail before anyone notices. Teams on the same stack and infrastructure recreate the same procedures. Meanwhile, a skill that worked for its original task can age as models change. Maintaining quality therefore requires testing against both the intended task and newer models, rather than treating the first successful run as permanent validation.
Discovery and ownership are linked. Touil invokes Backstage and the internal developer portal: a service catalog lets an engineer find who owns a microservice without asking around. A skills catalog needs the same property. Finding the procedure is useful; finding the person responsible for fixing and maintaining it makes reuse sustainable. Composition also needs coordination over task responsibilities and domains, because publishing several skills does not automatically make them work together.
Security has two separate mechanisms to address:
- Imported instructions and scripts: Public skills can contain prompt injection or insecure scripts. Because a skill can execute a script for a task, bringing one into the environment requires a checking pipeline for executable content as well as instructions.
- Sensitive business logic: Some internal skills encode knowledge that everyone in the organization should not access. Access control must govern who can retrieve and use those procedures.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Build the catalog, then assign people to maintain its domains
Adoption starts locally. Individuals need an agreed, structured way to create, test, improve, and use skills. Sharing with the team comes next, allowing people working on the same products and stack to refine the procedure together. The centralized platform then makes that work available beyond the originating team.
The platform needs several capabilities with distinct jobs:
- Searchable catalog and metadata: Help agents and engineers discover relevant skills. Touil proposes an MCP interface for catalog search and a CLI for pulling skills into an IDE or sandbox.
- Dependencies: Record relationships between skills so the platform can account for what a procedure needs.
- Versioning and lifecycle: Identify current releases and let an agent detect and retrieve an updated skill while building a feature.
- Access control: Make it possible to understand and govern who accesses which skills.
- Evaluation and observability: Keep quality assessment and operational visibility alongside distribution.
The hard question arrives after those capabilities exist: who governs them? Architecture leads, infrastructure leads, and cybersecurity leads need to own appropriate domains and ensure updates follow organizational policies. The precise assignment depends on the company’s structure. A platform can distribute a release; people must decide what belongs in it and what changes are acceptable. This is where, in Touil’s phrasing, technology stops solving the problem.
What makes reuse become shared improvement rather than one-way downloading? The circulation below connects a maintained catalog to local execution and contributions back. Teams pull shared skills, use them, and return improvements. Named domain owners provide the policy responsibility around that circulation.
Searchable metadata, dependencies, versions, access control, evaluation, and observability.
The catalog distributes procedures to local environments, while improvements return for shared use under human domain ownership.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make existing work visible before building it again
Touil illustrates the organizational effect with a simulation of fifteen teams over six months. Its inputs include engineers’ skill contributions, average daily utilization, duplication between teams, and quality and security ratios. It illustrates uneven productivity, cost, and implementation quality under different maintenance practices; it is a constructed scenario rather than measured evidence of a six-month organizational deployment.
The regulatory example explains the cost mechanism behind the scenario. Without a maintained procedure, a human repeatedly steers the agent toward the required rules—“vibe coding back and forth.” Each correction consumes time and tokens, and the outcome depends on that person’s knowledge. A reusable regulation skill supplies the procedure upfront, reducing the need to reconstruct it during every implementation.
The governed scenario changes what happens when the next engineer tries to create a skill. Once a shared skill is published, the coding harness can discover that it already exists and pull it. Discovery interrupts duplication at the moment it would otherwise begin. Teams gain common procedures even while their work remains different.
The same approach must extend to complete workflows. Skills are only one component. An engineer who needs to provision infrastructure should be able to retrieve a workflow with its required skills, run and test it, and return improvements to the centralized platform. That shares the arrangement of the work as well as the individual procedures it uses.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Evaluate skills before accelerating their evolution
The closing recommendations begin with a skills registry. Touil points to internal developer portals as a familiar place for centralizing this capability, alongside tools dedicated to skills. The practical requirement is a shared place to discover and maintain procedures; the product choice can follow the organization’s existing platform.
Evaluation remains an open design question. A useful starting point is linting or static checks against Anthropic’s skill best practices, looking for structural and invocation problems. A poorly structured skill may be hard for an agent to use; a skill that is not invoked appropriately cannot improve the intended task. These checks complement the earlier requirement to validate task performance and retest as models change.
Auto-evolving skills raise the stakes. A closed loop can keep changing procedures automatically, but Touil’s response to the promise is blunt: “So what?” The earlier maintenance problems still apply, now with a machine accelerating the changes. Ownership, evaluation, security checks, and organizational policies need to govern that loop. Otherwise, the organization is simply maintaining automatically changing skills without a way to keep their evolution aligned with its needs.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Related talks
- Building Great Agent Skills: The Missing Manual
A complementary next topic for designing the individual procedures that an organizational catalog will distribute.
- Don't Ship Skills Without Evals
Continue with the evaluation problem Touil identifies as essential to maintaining skill quality.
- Agentic Security: Permissions, Provenance, and the Agent Supply Chain
Explore the security questions raised by importing executable skills and controlling access to organizational knowledge.
Read the complete timestamped transcript
- 0:01
[music]
- 0:13
Thank you for joining me today. My name
- 0:16
is Imatel. I'm a distinguished engineer
- 0:18
at Quantum Black. And today talk I want
- 0:21
to really cover the uh how AI native
- 0:24
organizations runs on skills. But before
- 0:26
I really get started, um I just want to
- 0:28
do like a quick exercise to show of
- 0:30
hands. Can you raise your hand if you
- 0:32
have already created and using skills?
- 0:36
Amazing. Now can you keep your hand up
- 0:38
if you are using it and sharing it
- 0:40
within your teams?
- 0:42
Great. Now keep your hands up if you
- 0:45
have governed maintained skills across
- 0:48
your organization.
- 0:50
Amazing. I see a few hands. But that is
- 0:52
what this talk is about. Today I'm
- 0:54
really looking to break down why this is
- 0:56
really critical, how it's important and
- 0:58
how actually you can adopt it across
- 0:59
your organization. But before I get
- 1:01
started, what I really want to cover is
- 1:04
actually the agentic software stack. So
- 1:07
the agentic software stack has two
- 1:08
loops. So the first loop is all you
- 1:10
know, right? It's the code and agents or
- 1:11
the code agents harness right
- 1:14
at this you have some core components.
- 1:16
You will have your context manager, the
- 1:17
tool and MCPS, memories and states and
- 1:20
skills loader, right? But then there's
- 1:22
an outer loop which is your workflows,
- 1:25
right? Those has skills, sub aents, MCP
- 1:29
servers that you use and some hooks.
- 1:31
Sometimes you need them to have this
- 1:33
running properly. You will need some
- 1:35
enablement components at the bottom. So
- 1:37
what you would have is an environment
- 1:39
sandbox. You would have your MCP gateway
- 1:41
to manage and simplify all of the MCP
- 1:43
tools across your organization. a model
- 1:45
gateway to again manage and optimize for
- 1:48
all of your LMS either like you know
- 1:51
open source like running locally or or
- 1:54
or frontier models and also a graph
- 1:57
knowledge graph that abstracts your IT
- 2:00
core systems your codebase your skills
- 2:02
registry and at the end workflow
- 2:04
marketplace and then you have your
- 2:06
context layer right the context layer
- 2:08
will bring all of what is needed to get
- 2:10
the task done so this is the project
- 2:12
instruction think about it as the cloud
- 2:14
code MD file the agency file your tools
- 2:16
and MCP schema is actually to understand
- 2:18
which tool to use and when your memory
- 2:22
right conversation history with the end
- 2:24
user the human is in the loop and
- 2:26
finally the retrieved contents that's
- 2:28
you can pull from either like files your
- 2:30
codebase etc now what I want to really
- 2:33
focus on today is the workflow right and
- 2:36
I think this is where we kind of like
- 2:38
think um it's kind of a pretty simple um
- 2:41
uh workflow in day-to-day when you're
- 2:43
trying to actually create an end to-end
- 2:45
product software delivery life cycle in
- 2:47
the organization. In reality, we all
- 2:49
have seen the four steps. specify to
- 2:52
define what you want to build the design
- 2:53
or plan to plan you know what what are
- 2:56
you looking to build then you go to the
- 2:59
tasks you break it down into tasks and
- 3:01
then finally you start implementing
- 3:03
right I think this looks familiar I
- 3:05
think this is how most of our coding
- 3:07
agents actually are kind of shaped today
- 3:09
in reality that is not how it is
- 3:11
composed and at scale when you look at
- 3:13
the organization complexity
- 3:16
this is just one step in the journey
- 3:18
this is like building a product
- 3:19
increment
- 3:20
When you really look at the overall end
- 3:23
to end life cycle of something that a
- 3:25
business want to build to capture the
- 3:27
value out of it all the way to ship it
- 3:28
to the client, you will start first of
- 3:31
all by defining your product strategy.
- 3:32
What to build and how to build it,
- 3:34
right? You set the define the success
- 3:37
metrics. You identify and you break down
- 3:39
your plan road map for your products.
- 3:42
And to do this, you may need a lot of
- 3:43
insights, right? So then you do like
- 3:44
your market research. You do a
- 3:46
competitive analysis. You bring this as
- 3:48
an input with some customer interviews.
- 3:51
Then you go to the discovery side,
- 3:52
right? So then you start discovering.
- 3:54
Okay, now I understand what to build.
- 3:56
I'm going to break this down into like
- 3:57
some problem statements, find the
- 3:59
solution, validate the solution, and
- 4:01
then probably like experiment and then
- 4:03
create user stories. And before we start
- 4:06
building in reality actually need to
- 4:08
prepare your data, right? In some cases,
- 4:11
you will need actually to clean up your
- 4:13
data catalog that will support the build
- 4:15
of your products or maybe adjust some of
- 4:17
the endpoint connection and integration
- 4:19
to your core systems that will help you
- 4:21
actually build your products.
- 4:24
And this is where the data product
- 4:25
delivery comes. So you build your data
- 4:28
pipeline, you validate your data quality
- 4:30
and you put your catalog your
- 4:33
[clears throat] catalog data assets
- 4:34
ready for development. Then we kind of
- 4:37
like go back to the product increment.
- 4:39
That is where we will start. But then
- 4:41
building a product in every organization
- 4:43
that I have been uh serving for the past
- 4:45
18 years in my career. I can see that in
- 4:48
one organization you will find like
- 4:49
different DLC iss kind of like uh you
- 4:52
know scatterers across organization.
- 4:54
Some of it is actually for a mobile
- 4:55
application other is a different
- 4:57
department or different platform. Some
- 4:59
of it is internal platform that is for
- 5:01
your um employees. Other is actually
- 5:03
customerf facing. So it is not like a
- 5:05
one workflow that can actually build
- 5:06
anything you want for your organization.
- 5:09
When you figure out like what to build
- 5:10
and how to build it, you need to run it.
- 5:12
Then we come to the platform engineering
- 5:13
ops, right? That is again when you have
- 5:15
your provisioned infrastructure,
- 5:17
thinking about how you you build your
- 5:19
your infrastructure as code modules etc.
- 5:22
And then you launch your products and
- 5:24
the moment you launch it then you start
- 5:26
kind of like the journey of optimizing
- 5:28
the performance of your products trying
- 5:30
to look forward like any incidents to
- 5:32
resolve and then you start the loop
- 5:34
again. Right? So at scale when you look
- 5:37
at really building a digital platform
- 5:39
not like a very simple products that you
- 5:41
can you know solo build and deploy the
- 5:44
landscape is way complex that is
- 5:46
expected and what you're looking at here
- 5:47
is literally like probably 10 20% of
- 5:49
what is it and is really different from
- 5:51
organization to organization now going
- 5:53
back to the uh the stack right when you
- 5:57
look at the workflows there's like four
- 5:58
core components the first one is hooks
- 6:02
MCP servers and sub aents those are kind
- 6:04
of
- 6:05
given but it doesn't really bring the
- 6:07
right kind of like structured value to
- 6:09
your workflows right that's why scales
- 6:11
is one of the critical components in
- 6:13
your workflows hooks basically what it
- 6:15
does it just kind of like pre kind of
- 6:17
like trigon events to to do something
- 6:19
along your workflow the MCP service we
- 6:22
all know that you know you may need an
- 6:23
MCP tool but tell me like who actually
- 6:26
build a lot of MCPS we just use MCPS
- 6:28
tools that is actually provided by the
- 6:31
tool that we used to use before right so
- 6:33
we don't really own
- 6:34
The sub agents is just to minimize the
- 6:36
context window. We just delegate to sub
- 6:38
agent to execute a specific task what is
- 6:40
needed. So at the end of the day you
- 6:42
will find all of your knowhow is
- 6:44
actually at the skills level. And if you
- 6:46
don't have the right structure of your
- 6:47
skills then you're not really having a
- 6:49
deterministic workflow. And one thing to
- 6:52
mention is workflows think about them as
- 6:55
harness blueprints that actually shape
- 6:56
the behavior of your coding harness for
- 6:59
example in the runtime.
- 7:03
Now looking at the rise of skills
- 7:05
adoption right so just eight months ago
- 7:08
entropic published the first article
- 7:10
about skills right two month later we
- 7:13
had this standard an open standard that
- 7:15
is adopted and starts a lot of agent
- 7:18
harnesses starts adopting this new
- 7:20
standard [snorts] around February um
- 7:22
this year we have seen most of the
- 7:25
actually agents adopted this even though
- 7:27
you don't see them right if you pay
- 7:29
attention when the the uh the agent is
- 7:31
thinking And you can see that he's
- 7:32
pulling skills um as he's going and
- 7:35
doing the task. The other implication
- 7:37
here that you need to pay attention to
- 7:39
is actually the number of skills
- 7:40
created. Right? This is just like
- 7:41
literally a snapshot that I did across
- 7:43
some public GitHub repost, some public
- 7:45
skills registries, right? There's way
- 7:47
more than this publicly and within your
- 7:50
organizations, right? So the de the
- 7:52
creation of skills and the demand is
- 7:54
rising. But we need to understand why
- 7:58
in the latest skills bench comparing the
- 8:02
latest models against like you know
- 8:04
running the same task against auto
- 8:06
engineering and cyber security without
- 8:08
skills it did well right because that's
- 8:10
what expected and it's going to continue
- 8:12
to be improving day after day right but
- 8:15
then when we applied skills that's our
- 8:17
more deterministics the outcome was
- 8:19
clearly higher than than That's it.
- 8:24
Now if you think about now like how like
- 8:27
the anatomy of skills and how you
- 8:28
actually design and implement skills in
- 8:29
your organization. This is not like a
- 8:31
new problem that we are trying to solve
- 8:33
here, right? We have solved this with
- 8:34
the microser kind of movements, right?
- 8:36
So your microservices need to have all
- 8:38
of these design principles, right? This
- 8:40
is like a software kind of like you know
- 8:41
problem that we use to solve is is
- 8:43
similar. So skills need to be reusable,
- 8:45
right? Need to be modular, need to be
- 8:48
disco like you can discover your skills.
- 8:50
So if you are sitting in one team and
- 8:52
you need the skills you actually can
- 8:53
automatically discover and capture the
- 8:54
skills it's portable that you can
- 8:57
actually use skills across workflows but
- 8:59
also you can use skills across harnesses
- 9:00
again everyone adopted the same
- 9:02
standard. So if I'm having a skill on
- 9:04
cloud code and I want to move it to
- 9:05
cursor, it's going to just work.
- 9:08
Specialized skills that is where the
- 9:09
value you should not build like a one
- 9:11
skill like a monom or again it should be
- 9:13
specialized to define one tasks
- 9:15
specifically. Composible uh skills need
- 9:18
to be designed in a way that's actually
- 9:19
can compose. So you don't have
- 9:20
duplication across your skills when
- 9:22
you're trying to run them conflicts each
- 9:24
other. Consistence that is actually what
- 9:27
one of the key items of skills is
- 9:28
actually consistency and deterministic.
- 9:30
And finally cost efficience and for cost
- 9:32
I can go for another hour talk but it's
- 9:35
basically the skills comes to solve a
- 9:37
key problem around the context window
- 9:39
right is actually putting with the
- 9:41
disclosure pro progressive disclosure
- 9:43
pattern the right skills the right
- 9:45
amount of skills in the right time to
- 9:46
solve the right problem and that's
- 9:48
reduce the token usage. This define a
- 9:51
new unit right that makes your knowhow
- 9:53
in your organization executable portable
- 9:56
and cheap.
- 9:58
On the right hand side, you just a very
- 9:59
simple example of data retention policy.
- 10:01
Right? When it came to regulation, you
- 10:04
need to understand and make sure to
- 10:05
instruct your agents while you
- 10:07
manipulating. For example, um your
- 10:09
customer data, you should make sure that
- 10:12
this data manipulates according to the
- 10:14
regulation, right? And that bring me to
- 10:17
the next example.
- 10:20
On the left hand side, you can see that
- 10:21
there's like a think about like a a
- 10:24
catalog of skills. And in the right hand
- 10:26
side is your harness and the output is
- 10:28
on the right hand side. Right? So the
- 10:30
composible skills at the regulation
- 10:32
level. You have the skill that I just
- 10:34
showed earlier which is the retention
- 10:36
policy. But you will need disclosure
- 10:38
standards. You will need the GDPR rules
- 10:40
that need to be respected. You will need
- 10:42
the fill in templates. Right? So all of
- 10:44
this kind of think about it is like
- 10:45
defining how any data any feature that
- 10:48
is built across your web, mobile um like
- 10:51
you know different applications across
- 10:53
your organizations is really respecting
- 10:55
these rules and this gets pulled
- 10:57
automatically by on the runtime by the
- 10:59
regulatory disclosure review workflow
- 11:02
and the outcome is expected is
- 11:04
deterministic. You have an audit uh
- 11:07
audit reports that you can actually
- 11:09
store. you have um specific uh
- 11:12
identification of if there's anything to
- 11:14
improve and that this kind of like loop
- 11:15
back to improve your your codebase.
- 11:21
However, if we don't govern skills, we
- 11:23
will start creating a new class of
- 11:25
technical depths, right? First of all,
- 11:28
you will find out that you are having a
- 11:29
lot of duplication in your organization.
- 11:31
So if teams are not collaborating and
- 11:33
everyone is think about it using the
- 11:35
same technology stack the same
- 11:37
infrastructure you're for sure building
- 11:39
the same skills over and over again
- 11:41
without sharing them quality. If you
- 11:44
don't test and make sure that you're
- 11:45
maintaining and you're validating your
- 11:46
skills not against your task but also
- 11:49
against the latest models that comes
- 11:50
right then the quality starts
- 11:52
degragating over time.
- 11:56
You should be able to discover your
- 11:57
skills in in reality without a
- 11:59
governance you cannot really discover
- 12:01
it. Think about it as the the backstage
- 12:03
the IDP right it's come to solve a
- 12:05
problem where okay I need to understand
- 12:07
who owns this microser back in the days
- 12:09
right I don't need to talk to anyone I
- 12:10
just need to tap into the service
- 12:12
catalog and immediately find who
- 12:13
actually own it that is brings ownership
- 12:16
part if you don't have an owner then no
- 12:18
one will be able to maintain scale those
- 12:21
skills
- 12:22
composibility is not something that
- 12:24
comes by default you need to have a
- 12:26
governance way to to align what to build
- 12:28
and how to design it is think about the
- 12:30
domain driven approach that we have been
- 12:31
taken also for for many years, right? Is
- 12:33
similar to how you shape your skills
- 12:35
catalog.
- 12:37
Security again some of uh all of us like
- 12:40
we experiment with the public skills,
- 12:43
right? But when you think about it, some
- 12:45
skills may have some prompts injection
- 12:46
and skills actually does have scripts
- 12:48
because that is the deterministic part
- 12:50
of it because it can run a specific
- 12:51
script for a specific task. So if you
- 12:54
don't have again a pipeline that check
- 12:55
your security, you may be pulling
- 12:57
something that is insecure
- 12:59
and permissions. Not every skills is
- 13:01
actually something that anyone in the
- 13:03
organization should access. Some skills
- 13:04
may have some business logic that is
- 13:07
very sensitive, right? So the access
- 13:09
control is is also crucial at this
- 13:11
stage.
- 13:13
Now how to bring this to your
- 13:14
organization? First of all, you need to
- 13:16
allow at the individual level to create,
- 13:19
test, improve and use those skills.
- 13:22
Again, it shouldn't be random. It should
- 13:24
be structured way. There's a different
- 13:25
tools out there. You just need to decide
- 13:27
which tool actually you want to agree on
- 13:28
and you use that mechanism at the
- 13:30
individual level. The moment you create
- 13:32
a skills, you need to be sharing it with
- 13:33
your team, right? Your team starts
- 13:36
collab collaborating to improve the
- 13:38
skills and think about it. You build the
- 13:39
same technology stack, building the same
- 13:41
products. So, it's going to evolve
- 13:42
really quickly. But then you move on to
- 13:44
a very critical point which is the
- 13:47
centralized platform. That is where all
- 13:50
of what I've been covering so far come
- 13:52
to play. You need a centralized platform
- 13:54
that have a catalog with metadata in it
- 13:57
that's actually can discover skills and
- 13:59
could be searchable. You can have an MCP
- 14:00
that's actually plugged to this cat
- 14:02
catalog, search for the skill and a CLI
- 14:04
to pull the skills back to your either
- 14:06
your IDE if you're locally or to your um
- 14:09
sandbox in your factory. Then you have
- 14:11
the dependencies. So you need to
- 14:12
understand the dependencies between the
- 14:14
skills as well. You have the versioning
- 14:16
uh and and life cycle. So you understand
- 14:19
which version of the skills is actually
- 14:20
the latest. Um and a very good example
- 14:23
when I'm using for example building a
- 14:25
functionality I can the agents
- 14:27
automatically up capture that there is a
- 14:29
latest version of the skill and pull it
- 14:31
right. So this versioning help also to
- 14:33
pull the the the right latest changes
- 14:35
from the skills registry access control
- 14:38
again as I said if you don't know who is
- 14:39
accessing what that is a huge uh uh gap
- 14:43
and finally evaluation observ
- 14:44
observability and then all of this is
- 14:47
actually play around a governance and
- 14:49
this is where technology stop solving
- 14:51
the problem right so you figure out all
- 14:52
of this all good now who's going to
- 14:54
govern this and that is where dep you
- 14:58
know it really depend how your
- 14:59
organization is structured today that is
- 15:01
where you should have your architects,
- 15:02
your engineer leads, infra leads etc and
- 15:05
cyber leads actually sitting down owning
- 15:08
part of those domains and making sure
- 15:09
that the skills we need to get updated
- 15:11
is actually according to the policies
- 15:12
you want to adhere within your
- 15:14
organization and drive this change.
- 15:17
And finally when you get this right what
- 15:19
you will have you would have at the
- 15:20
organization level all of your teams
- 15:23
pulling from one centralized place high
- 15:25
quality skills executing them and
- 15:27
pulling them back um to to the
- 15:29
centralized platform if it is improved.
- 15:32
Now what I want to bring this because
- 15:33
it's a little bit of a inclear view. So
- 15:35
what I created I created a simulation
- 15:37
right. So think about this. This is your
- 15:40
organization today, right? And what I
- 15:42
have here, I have um just a random
- 15:45
teams. I have 15 teams created um five
- 15:48
to 12 like per team. Uh you have skills
- 15:52
per engineers contribution. You have the
- 15:54
average skills utilization kind of like
- 15:56
on average like how much time uh skills
- 15:58
are being pulled per day, the
- 16:00
duplication across the team kind of as a
- 16:02
ratio and the skills quality and
- 16:04
security ratio. Now if I run this across
- 16:07
six months what's really happening and
- 16:09
think about it this is already happening
- 16:11
within your organization is teams are
- 16:13
creating and using the skills right but
- 16:16
we don't have visibility and skills
- 16:18
again they're tightly coupled to your
- 16:21
productivity uplift if for example the
- 16:23
example that I shared earlier on the
- 16:25
regulation if we don't have a skill
- 16:27
about the regulation that is someone is
- 16:29
vip coding back and forth and trying to
- 16:30
figure out exactly how to steer the
- 16:32
agent to implement it properly right
- 16:34
that is burning more tokens from one
- 16:36
side costwise but also the productivity
- 16:38
is spending more time rather than giving
- 16:40
in one shot the right answer and the
- 16:42
quality and security is similar if you
- 16:44
don't have clear you know skills defined
- 16:46
and maintained you will have a low
- 16:49
quality in your implementation because
- 16:51
then it's up to the human to decide this
- 16:53
and different the m in the maturity from
- 16:55
team to team you can see the difference
- 16:57
and that is why for example if I look uh
- 16:59
just randomly at this this is like you
- 17:01
can see the productivity of this team is
- 17:03
a kind of a medium right If I look at
- 17:05
this one is a is a bit of like I don't
- 17:07
know it's low medium productivity
- 17:10
quality and security also medium but
- 17:12
when it came to the cost is really high.
- 17:15
Okay, now let's actually say okay how
- 17:18
this looks like if I governed all of the
- 17:21
my skills in my in my organizations.
- 17:24
What's going to happen is of course some
- 17:26
of them will split right and this is
- 17:27
reality isn't going to be perfect um as
- 17:30
we expect but at least what you will see
- 17:32
you will see actually some common ground
- 17:35
across all of your teams. The moment you
- 17:36
govern, you publish one skill, the next
- 17:39
engineer trying to build a new skill,
- 17:41
the coding agent harness will identify
- 17:43
this skill that is already available and
- 17:45
pull it. Right? So you you almost solve
- 17:47
all of the issues that that I covered
- 17:49
about the governance.
- 17:53
And one last point is when it came to
- 17:55
skills, um it's just one component of
- 17:58
your workflows as I said, right? So that
- 18:00
doesn't mean uh if you figure out skills
- 18:02
that says you're good. No, you need to
- 18:04
apply the same kind of like approach and
- 18:07
solution for your whole workflows. And
- 18:09
you may think to apply this again if you
- 18:11
think about it like if you have a a
- 18:13
centralized platform that have all of
- 18:15
your workflows, right? From one side
- 18:16
you're centralizing the workflows which
- 18:18
is also having the skills but also if
- 18:20
the next engineer came and want to I
- 18:22
don't know like provision infrastructure
- 18:24
they can tap into a workflow and build
- 18:26
that workflow with the required skills
- 18:28
and run it and test it again. And if
- 18:30
it's something you need to be improved
- 18:31
in the workflow, you can easily push it
- 18:33
back um to the centralized platform for
- 18:36
your organization to use.
- 18:40
[snorts] Now before I wrap up, um what I
- 18:44
want to leave you with is this is just
- 18:45
the start the beginning. You see like
- 18:47
this just we're talking about six to
- 18:48
eight months. What's coming next and I
- 18:51
would invite you to already explore is
- 18:53
skills registry, right? You should have
- 18:55
one if not already. And the good news is
- 18:57
all of the players that's been solving
- 18:59
the IDP problem like um internal
- 19:01
developer portal they already start
- 19:03
centralizing this capability right so if
- 19:06
you don't have it today maybe in a
- 19:07
couple of months you will see it coming
- 19:08
but also there's a lot of tools that's
- 19:10
actually solving this specific problem
- 19:13
second is skills eval evaluation there's
- 19:15
still kind of like a discussion on what
- 19:17
is the right approach to to to you know
- 19:19
to to evaluate skills the easy thing
- 19:21
that I found so far very valuable is
- 19:23
actually test like you you l static test
- 19:26
your or evaluate your skills against the
- 19:28
entropic best practices. Right? If the
- 19:30
skill is not invoked properly, if the
- 19:32
skill is not structured properly,
- 19:34
there's a high chance that it's not
- 19:35
going to be high quality. And finally is
- 19:37
auto evolving. And again, this is what
- 19:39
everyone kind of like is is the next
- 19:41
hype right now. Like yeah, I can create
- 19:43
like a an a closed loop that can evolve
- 19:45
automatically my skills. So what right?
- 19:48
If you automatically start this machine,
- 19:50
the impact will be way more than it is
- 19:53
today because what I shared earlier is
- 19:55
going to be just maintaining auto
- 19:57
evolving skills without that governance
- 20:00
in place that actually put the
- 20:01
guardrails for your organization.
- 20:03
And at this point I would leave you
- 20:05
here. Thank you so much for your
- 20:07
listening and looking forward if you
- 20:08
have any question I will be in the
- 20:10
leadership lounge. Feel free to grab me.
- 20:12
Thank you so much. [applause]
- 20:28
>> [music]