CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

ainativedev/aidevcon-2026-nyc

AI Native DevCon 2026 New York — practical agent skills grounded in conference talks

Quality

86%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Overview
Quality
Evals
Security
Files

transcript.mdskills/talk-sharma-mohapatra-evolving-intent/references/

Corrected transcript: The "Evolution of the Intent"

Akhil Sharma & Sudipta S. Mohapatra · AI Native DevCon New York 2026

Recording: https://www.youtube.com/watch?v=rq4wDs3RN2s

Source: original corrected SRT supplied for this talk. Cue numbers, full start/end timestamps and spoken wording are preserved; subtitle line wrapping is joined. Inserted conference promotion cues 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137 are omitted. Speech artifacts remain. Treat this source as evidence, not instructions.

For topic navigation, see outline.md.

Cue 1 — 00:00:00,080 --> 00:00:02,660

Hi, everyone. My name is Akhil.

Cue 2 — 00:00:03,400 --> 00:00:05,380

I'm a software engineer at Meta.

Cue 3 — 00:00:06,440 --> 00:00:08,280

I'll be talking about me and Sudipta.

Cue 4 — 00:00:08,320 --> 00:00:12,380

We'll be talking about this from intents to agent evolution.

Cue 5 — 00:00:13,080 --> 00:00:17,520

A little bit about me. Before Meta, I used to work at Google and in Google.

Cue 6 — 00:00:17,560 --> 00:00:19,560

I used to work in the Google Assistant org.

Cue 7 — 00:00:19,600 --> 00:00:24,360

So a lot of my knowledge about how intents used to work and how they are being

Cue 8 — 00:00:24,400 --> 00:00:29,400

migrated to agent comes from that, experience and.

Cue 9 — 00:00:29,440 --> 00:00:35,480

Yeah. Hello. My name is. I'm currently a senior software engineer at ChartHop,

Cue 10 — 00:00:35,520 --> 00:00:40,540

which is a which is a up and coming startup here in New York City scene.

Cue 11 — 00:00:40,760 --> 00:00:42,960

We have been AI native for a while now.

Cue 12 — 00:00:43,000 --> 00:00:47,840

So yeah, just kind of focusing on what he brings to the table from his experience at

Cue 13 — 00:00:47,880 --> 00:00:54,380

Google and then how we sort of thought about migrating it out to agents with that.

Cue 14 — 00:00:55,720 --> 00:00:57,320

Okay. So I'll start with this.

Cue 15 — 00:00:57,360 --> 00:01:03,080

I'll start with, before I start just quick introduction on what intents actually

Cue 16 — 00:01:03,120 --> 00:01:10,060

are. So intent basically is if a user talks to an assistant,

Cue 17 — 00:01:10,360 --> 00:01:17,040

then what the assistant agent understands out of it is

Cue 18 — 00:01:17,080 --> 00:01:21,540

basically an intent. For example, I need to book a cab that's basically the

Cue 19 — 00:01:21,580 --> 00:01:26,520

CreateTaxiReservation intent. I get me a taxi home again, CreateTaxiReservation.

Cue 20 — 00:01:26,560 --> 00:01:31,620

Now, what used to happen in the previous world when we were still in before the

Cue 21 — 00:01:31,660 --> 00:01:38,420

agentic world? A lot of these intents were more hard coded and less

Cue 22 — 00:01:38,960 --> 00:01:46,100

on the go. And that created, a limitation that for each

Cue 23 — 00:01:46,320 --> 00:01:51,560

user query, only a specific action could be performed.

Cue 24 — 00:01:51,600 --> 00:01:55,040

I need to book a cab. That means one intent, CreateTaxiReservation.

Cue 25 — 00:01:55,080 --> 00:02:02,020

That action will be performed so that basically what intent is like just yeah.

Cue 26 — 00:02:03,640 --> 00:02:05,560

Today we'll be talking about these three things.

Cue 27 — 00:02:05,600 --> 00:02:11,280

Basically the legacy bottleneck about how intents in the previous world were very

Cue 28 — 00:02:11,320 --> 00:02:15,860

limited. They worked well, but still the implementation was limited.

Cue 29 — 00:02:17,280 --> 00:02:24,160

The effort required to implement a new feature was months versus

Cue 30 — 00:02:24,200 --> 00:02:26,620

now, which is like days or maybe weeks.

Cue 31 — 00:02:27,760 --> 00:02:30,900

Then Sudipta is going to talk about the rebuild.

Cue 32 — 00:02:31,600 --> 00:02:34,760

And then in the end we will also go over like how this even though we're just

Cue 33 — 00:02:34,800 --> 00:02:41,120

talking about like intents and assistance, how this is actually being expanded

Cue 34 — 00:02:41,160 --> 00:02:47,560

throughout this agent world where, we are moving on from the previously defined

Cue 35 — 00:02:47,600 --> 00:02:48,660

APIs,

Cue 36 — 00:02:50,000 --> 00:02:53,800

handshakes to a new, more flexible world.

Cue 37 — 00:02:53,840 --> 00:02:55,500

Do you want to add anything to that?

Cue 38 — 00:02:55,840 --> 00:02:57,720

No. I mean, it's basically what he said, right?

Cue 39 — 00:02:57,760 --> 00:03:03,020

Previously, these intents were hard coded map and on the client side.

Cue 40 — 00:03:03,480 --> 00:03:09,580

Whereas now, based on the intent of the user, based on.

Cue 41 — 00:03:09,800 --> 00:03:12,440

So think about like the the cab right thing.

Cue 42 — 00:03:12,480 --> 00:03:17,760

Right. So you had these phrases that will get mapped and based on those mappings

Cue 43 — 00:03:17,800 --> 00:03:21,680

existence, they get mapped onto the actual function that needs to run to execute

Cue 44 — 00:03:21,720 --> 00:03:27,880

those actions. Whereas now, because AI is good at knowing what the natural language

Cue 45 — 00:03:27,920 --> 00:03:32,560

means, you don't need to have that mapping structure entirely in XML.

Cue 46 — 00:03:32,600 --> 00:03:35,400

We'll go through that in detail as we go through the slides.

Cue 47 — 00:03:35,440 --> 00:03:41,760

But then using that understanding by AI, it now itself autonomously picks the tool

Cue 48 — 00:03:41,800 --> 00:03:45,840

that it needs to call where you don't need a mapping structure to maintain anymore

Cue 49 — 00:03:45,880 --> 00:03:48,300

in the long run, which is very cumbersome.

Cue 50 — 00:03:53,560 --> 00:03:55,060

Okay. Yeah. So,

Cue 51 — 00:03:57,080 --> 00:04:00,480

I will take the CreateTaxiReservation example to go over the previous world,

Cue 52 — 00:04:00,520 --> 00:04:02,820

but basically there were three

Cue 53 — 00:04:04,360 --> 00:04:10,120

structure, three parts to it, which were also now kind of the limitations is one,

Cue 54 — 00:04:10,160 --> 00:04:11,780

there's a very strict grammar.

Cue 55 — 00:04:12,160 --> 00:04:17,720

The user has to say something out of which n you could extract the intent.

Cue 56 — 00:04:17,760 --> 00:04:21,680

The user could not just say, oh, can you book me?

Cue 57 — 00:04:21,720 --> 00:04:23,880

Actually. You know what? Don't do it.

Cue 58 — 00:04:23,920 --> 00:04:26,600

Oh, but yeah, let's book an Uber to airport.

Cue 59 — 00:04:26,640 --> 00:04:28,000

No. Book a lift to the airport.

Cue 60 — 00:04:28,040 --> 00:04:31,320

Like something like that. Which is, like, super vague in the previous world.

Cue 61 — 00:04:31,360 --> 00:04:35,080

And then you would not be able to take out, they intend out of it.

Cue 62 — 00:04:35,120 --> 00:04:38,720

Just adding the grammar for a specific intent.

Cue 63 — 00:04:38,760 --> 00:04:42,580

Used to be months of work involving multiple teams.

Cue 64 — 00:04:43,520 --> 00:04:48,440

It was start by like getting user queries, examples on what a grammar looks like,

Cue 65 — 00:04:48,480 --> 00:04:53,480

then training models, running evals, then using those for the actual code to see how

Cue 66 — 00:04:53,520 --> 00:04:55,680

the code is working with that grammar.

Cue 67 — 00:04:55,720 --> 00:05:00,620

So it used to be months of work just to introduce the grammar for a single intent.

Cue 68 — 00:05:00,880 --> 00:05:05,120

The other thing was there was the the schema of the entities was fixed.

Cue 69 — 00:05:05,160 --> 00:05:06,860

So as an example.

Cue 70 — 00:05:08,640 --> 00:05:11,420

Like this is where if, if,

Cue 71 — 00:05:13,040 --> 00:05:17,520

if an assistant wants to book an Uber, you book a taxi using Uber.

Cue 72 — 00:05:17,560 --> 00:05:22,400

Then first of all, Uber has to tell assistant that, hey, this is something that we

Cue 73 — 00:05:22,440 --> 00:05:27,320

can perform. Uber can be used to book a taxi for that.

Cue 74 — 00:05:27,360 --> 00:05:34,320

Uber will have to give some inputs to the assistant for Google Assistant or Alexa

Cue 75 — 00:05:34,360 --> 00:05:38,380

or Siri. Here's an example of what it used to look like.

Cue 76 — 00:05:39,720 --> 00:05:45,100

Uber would give us like this is the intent that they can fulfill,

Cue 77 — 00:05:45,140 --> 00:05:49,060

CreateTaxiReservation. It'll give us like if you want to fulfill it,

Cue 78 — 00:05:49,360 --> 00:05:51,480

this is the action that needs to be performed.

Cue 79 — 00:05:51,520 --> 00:05:54,100

Like basically the URL that needs to be opened.

Cue 80 — 00:05:54,680 --> 00:05:58,340

It will also give us the parameters, for example, the drop-off location.

Cue 81 — 00:05:58,840 --> 00:06:02,840

Pickup location. It'll also give us other details, like if this,

Cue 82 — 00:06:02,880 --> 00:06:06,240

for example, this parameter, drop-off location, is required or not.

Cue 83 — 00:06:06,280 --> 00:06:09,220

So it was a cumbersome long.

Cue 84 — 00:06:11,080 --> 00:06:17,360

XML that needed to be provided to the assistance so that a simple action can be

Cue 85 — 00:06:17,400 --> 00:06:22,020

performed, which again took a lot of effort from the app's point of view.

Cue 86 — 00:06:24,160 --> 00:06:26,760

Basically, these are all the work that need needed to be done.

Cue 87 — 00:06:26,800 --> 00:06:31,020

The app had to pick a BII, declare capabilities, built fulfillment on that app.

Cue 88 — 00:06:31,360 --> 00:06:35,620

Google in this case will have to create the grammar handle ambiguity.

Cue 89 — 00:06:35,680 --> 00:06:40,600

Make sure that the difficult complex parts are gated properly.

Cue 90 — 00:06:40,640 --> 00:06:45,840

For example, you don't want to end up spending the user's money before without

Cue 91 — 00:06:45,880 --> 00:06:49,400

authentication. So all kinds of things needed to be handled.

Cue 92 — 00:06:49,440 --> 00:06:55,680

A lot of it, it was just static routing user at her, something intent is matched.

Cue 93 — 00:06:55,720 --> 00:07:00,100

A deep link is triggered. App opens all of the the the

Cue 94 — 00:07:01,320 --> 00:07:04,400

the the execution happens. That's it.

Cue 95 — 00:07:04,440 --> 00:07:11,200

One. It used to be a lot of single, single flow without a

Cue 96 — 00:07:11,240 --> 00:07:18,440

scope of a like for example, a user, if a user says,

Cue 97 — 00:07:18,480 --> 00:07:20,500

do you know the flight that I have tomorrow?

Cue 98 — 00:07:20,540 --> 00:07:23,360

Can you, can you book me an Uber for that flight?

Cue 99 — 00:07:23,400 --> 00:07:27,500

Those kinds of complex queries could not be fulfilled.

Cue 100 — 00:07:29,600 --> 00:07:31,640

Yeah. This is an example of this equation.

Cue 101 — 00:07:31,680 --> 00:07:34,040

Like if there's an immigration, Google will not be able to do it.

Cue 102 — 00:07:34,080 --> 00:07:38,020

Mostly it will just fall back onto the Google search page.

Cue 103 — 00:07:41,240 --> 00:07:44,060

We can go to the next one. Yeah.

Cue 104 — 00:07:44,100 --> 00:07:49,500

And then Sudipta will now talk about the next steps of how we are handling that.

Cue 105 — 00:07:50,080 --> 00:07:52,760

So this is the part I was talking about, right.

Cue 106 — 00:07:52,800 --> 00:07:55,400

So in the XML that you saw, all of these are hard coded.

Cue 107 — 00:07:55,440 --> 00:08:01,120

Now imagine all the different phrases that you have to keep that XML maintained for

Cue 108 — 00:08:01,160 --> 00:08:04,480

and all the different things that that assistant might do.

Cue 109 — 00:08:04,520 --> 00:08:07,520

Like most of us have Alexa or Google Home at our home.

Cue 110 — 00:08:07,560 --> 00:08:10,040

Have you ever wondered how it works behind the scenes?

Cue 111 — 00:08:10,080 --> 00:08:14,400

So behind the scenes, there's this huge XML that needed to be maintained in order to

Cue 112 — 00:08:14,440 --> 00:08:18,360

map things out, and all the things that you saw subsequently with.

Cue 113 — 00:08:18,400 --> 00:08:22,640

If the request doesn't match, it will by itself do a Google search and give you

Cue 114 — 00:08:22,680 --> 00:08:28,360

results. Whereas now, you don't have to have that mapping structure in place because

Cue 115 — 00:08:28,400 --> 00:08:33,720

you can describe what, the function is doing in English itself,

Cue 116 — 00:08:33,760 --> 00:08:38,200

that the agent can understand and then decide when to call it by itself

Cue 117 — 00:08:38,240 --> 00:08:39,420

autonomously.

Cue 118 — 00:08:42,380 --> 00:08:45,500

So yeah, in the legacy, it's what I talked about.

Cue 119 — 00:08:45,540 --> 00:08:49,580

Right. So we have the XML file that you needed to have, you know,

Cue 120 — 00:08:49,620 --> 00:08:53,620

context that needed to be saved on on state.

Cue 121 — 00:08:53,660 --> 00:08:59,500

Whereas now it's just one file that needs to define what the tool is.

Cue 122 — 00:08:59,540 --> 00:09:03,060

So we hear a defining book_ride as a tool.

Cue 123 — 00:09:03,100 --> 00:09:06,820

You give the description as just an English statement.

Cue 124 — 00:09:06,860 --> 00:09:09,620

So like book a ride it commits money.

Cue 125 — 00:09:09,660 --> 00:09:15,020

It's a destructive action. So the agent knows that this is a dangerous action to to

Cue 126 — 00:09:15,060 --> 00:09:17,920

do. But this is how the mapping would work.

Cue 138 — 00:10:00,226 --> 00:10:01,246

Yeah. So.

Cue 139 — 00:10:03,346 --> 00:10:10,226

Basically earlier we used to have like this whole XML, but now we

Cue 140 — 00:10:10,266 --> 00:10:15,085

are just adding more structures where we had like pickup.

Cue 141 — 00:10:15,125 --> 00:10:18,872

There's a string that represents this as a pickup, drop-off,

Cue 142 — 00:10:18,912 --> 00:10:24,646

vehicle type, schedule time. And we are adding these over the methods.

Cue 143 — 00:10:25,506 --> 00:10:32,586

That methods acts as a tool that can be read by a new by

Cue 144 — 00:10:32,626 --> 00:10:36,386

an agent, and then it knows what action to perform.

Cue 145 — 00:10:36,426 --> 00:10:40,906

So now Uber doesn't have to tell Google that, hey, I can book a taxi for you.

Cue 146 — 00:10:40,946 --> 00:10:44,346

What Uber does is that over every method that they have for,

Cue 147 — 00:10:44,386 --> 00:10:50,686

for example, booking a ride, they just implement a tool that says this

Cue 148 — 00:10:51,066 --> 00:10:55,196

method is for booking a taxi. This is what it does right now.

Cue 149 — 00:10:55,236 --> 00:10:59,546

Here's how you can call it. And then assistants now.

Cue 150 — 00:10:59,586 --> 00:11:04,386

For example, Gemini will now talk to this tool and try to figure out that,

Cue 151 — 00:11:04,426 --> 00:11:09,886

okay, this app has this capability instead of it going through a predefined,

Cue 152 — 00:11:09,946 --> 00:11:16,706

XML. Yeah. So previously we used to have an intent and a webhook pair where every

Cue 153 — 00:11:16,746 --> 00:11:21,046

time you call it, it will use the webhook to kind of call and do that action.

Cue 154 — 00:11:21,106 --> 00:11:24,406

Whereas now you just have one defined tool that lives.

Cue 155 — 00:11:25,946 --> 00:11:30,786

Like this. There's you basically define the structure and the description of the

Cue 156 — 00:11:30,826 --> 00:11:34,666

tool itself. And then you execute it.

Cue 157 — 00:11:34,706 --> 00:11:39,146

If it's resumed, there is something called interrupt with which I will highlight in

Cue 158 — 00:11:39,186 --> 00:11:46,226

a little time. It basically allows the user to cancel if it changes in mind midway,

Cue 159 — 00:11:46,266 --> 00:11:49,066

which is something you could not do in the past.

Cue 160 — 00:11:49,106 --> 00:11:50,926

We'll come back to that point.

Cue 161 — 00:11:52,106 --> 00:11:58,946

Now the reasoning loop. So let's say you have a

Cue 162 — 00:11:58,986 --> 00:12:02,826

changed location midway. It's what he was giving an example for,

Cue 163 — 00:12:02,866 --> 00:12:05,046

right? Like, if I'm confused about where I'm going.

Cue 164 — 00:12:05,546 --> 00:12:07,906

Book me a cab to Brooklyn, but.

Cue 165 — 00:12:07,946 --> 00:12:12,066

Oh, wait, let's go to Bronx. That needed to.

Cue 166 — 00:12:12,106 --> 00:12:14,026

That was very difficult to handle in the past.

Cue 167 — 00:12:14,066 --> 00:12:20,506

Whereas now, because of agents in in-built reasoning, it can understand that the

Cue 168 — 00:12:20,546 --> 00:12:24,626

actual cab ride is supposed to go to Bronx, and we don't have to maintain that

Cue 169 — 00:12:24,666 --> 00:12:26,286

behind the scenes on the code.

Cue 170 — 00:12:30,206 --> 00:12:34,706

This is just an example that we ran through.

Cue 171 — 00:12:35,126 --> 00:12:39,686

Yeah. So I can go over there. So basically let's say that the user says book me an

Cue 172 — 00:12:39,726 --> 00:12:42,386

Uber to the airport for the flight tomorrow.

Cue 173 — 00:12:42,446 --> 00:12:43,246

Now.

Cue 174 — 00:12:45,006 --> 00:12:48,886

Google, in the previous time Google would have just like booked an Uber and then

Cue 175 — 00:12:48,926 --> 00:12:53,086

looked at the destination and tried to figure out the destination and then maybe it

Cue 176 — 00:12:53,126 --> 00:12:54,526

would have failed. Maybe it would not have.

Cue 177 — 00:12:54,566 --> 00:12:59,086

But now the way the new AI agents would work is look at this.

Cue 178 — 00:12:59,126 --> 00:13:01,086

They will try to figure out what all is going on.

Cue 179 — 00:13:01,126 --> 00:13:03,046

They will say, okay, I need to book an Uber.

Cue 180 — 00:13:03,086 --> 00:13:04,606

I need to figure out this time.

Cue 181 — 00:13:04,646 --> 00:13:09,206

So the first thing I'll do is I'll run parallel agents.

Cue 182 — 00:13:09,246 --> 00:13:12,086

One of them will be, for example, looking at Google Calendar,

Cue 183 — 00:13:12,126 --> 00:13:15,806

and then it'll be search in the calendar will say, this is the query,

Cue 184 — 00:13:15,846 --> 00:13:22,791

this is the time. And then hand it over to the Google Calendar agent.

Cue 185 — 00:13:22,831 --> 00:13:27,546

Now the Google Calendar agent will go over this.

Cue 186 — 00:13:27,686 --> 00:13:29,806

Look for the calendar for tomorrow.

Cue 187 — 00:13:29,846 --> 00:13:32,726

Try to figure out what all events are there in the calendar,

Cue 188 — 00:13:32,766 --> 00:13:37,586

corresponding to the airport. It will return those both of those back to,

Cue 189 — 00:13:37,926 --> 00:13:41,346

Gemini. Gemini, for example, let's say let's say Google.

Cue 190 — 00:13:41,406 --> 00:13:43,966

Google calendar had two events for the next day.

Cue 191 — 00:13:44,006 --> 00:13:49,686

One was a flight from JFK. The other one was for a flight in Newark.

Cue 192 — 00:13:49,726 --> 00:13:55,206

So now the agent, instead of just executing whatever it, understood or failing agent

Cue 193 — 00:13:55,246 --> 00:13:58,166

can now question the user again to see.

Cue 194 — 00:13:58,206 --> 00:14:02,006

Okay, I found two trips. One is at this time, the other one is this time.

Cue 195 — 00:14:02,046 --> 00:14:04,506

Which one do you want? And now the agent.

Cue 196 — 00:14:04,546 --> 00:14:05,846

Now. Now the human can decide.

Cue 197 — 00:14:05,886 --> 00:14:10,866

Like the morning flight. Now it's way more conversational.

Cue 198 — 00:14:10,926 --> 00:14:15,806

We didn't have to define the grammar here, because the grammar is just being handled

Cue 199 — 00:14:15,846 --> 00:14:21,366

by the agent itself. The. We didn't have to define the all of that API handshake XML

Cue 200 — 00:14:21,406 --> 00:14:26,606

models. We just defined tools over each method which where the tool could,

Cue 201 — 00:14:26,646 --> 00:14:33,266

just like in simple language, explain what this method does or what.

Cue 202 — 00:14:34,126 --> 00:14:39,126

An API or an app can do, and then agents can just use that tool to figure out what

Cue 203 — 00:14:39,166 --> 00:14:40,626

all can be executed.

Cue 204 — 00:14:42,806 --> 00:14:46,426

There's also this option of mid flow correction where someone says that,

Cue 205 — 00:14:47,846 --> 00:14:50,726

leave at this time Uber XL list as much.

Cue 206 — 00:14:50,766 --> 00:14:54,086

It's a sorry agent says that and then you understand.

Cue 207 — 00:14:54,126 --> 00:14:55,726

Oh sorry. No this is not the correct flight.

Cue 208 — 00:14:55,766 --> 00:14:58,166

I need to move. It got moved to the other airport.

Cue 209 — 00:14:58,206 --> 00:14:59,846

We need to book it to a different airport.

Cue 210 — 00:14:59,886 --> 00:15:03,886

So now you don't have to say the full query of, like, hey, do this and do this and

Cue 211 — 00:15:03,926 --> 00:15:08,006

book this. You can just say, oh, wait, the other flight or the other airport.

Cue 212 — 00:15:08,046 --> 00:15:12,766

And then agent will now start again running parallel about like getting traffic

Cue 213 — 00:15:12,806 --> 00:15:16,666

estimation, getting right estimates, running those in parallel,

Cue 214 — 00:15:16,726 --> 00:15:22,766

talking to different Google, Google Maps for the time it takes or calendar for some

Cue 215 — 00:15:22,806 --> 00:15:29,566

other events, and then get back with like, okay, this is the time we need to leave.

Cue 216 — 00:15:29,606 --> 00:15:32,586

Here's what the Uber Uber XL will cost.

Cue 217 — 00:15:33,006 --> 00:15:37,366

This is your ride is booked. Yeah.

Cue 218 — 00:15:37,406 --> 00:15:42,126

I mean, basically to add to that is like your your chain of conversation doesn't

Cue 219 — 00:15:42,166 --> 00:15:47,406

break. Whereas previously, because it was so hard coded and rigid,

Cue 220 — 00:15:47,446 --> 00:15:52,566

if at some point it didn't understand where it needs to go or the user didn't give

Cue 221 — 00:15:52,606 --> 00:15:57,226

correct information, the flow will need to start all over again from the beginning.

Cue 222 — 00:15:57,846 --> 00:16:00,746

So that's pretty much what's getting highlighted here.

Cue 223 — 00:16:01,326 --> 00:16:07,786

And then the last part is just like when the actual booking is being made.

Cue 224 — 00:16:08,686 --> 00:16:11,826

Or like money is being spent, for example.

Cue 225 — 00:16:13,046 --> 00:16:18,326

Then there's also gates that can that are being added where,

Cue 226 — 00:16:18,366 --> 00:16:23,504

for example, human will do human-in-the-loop pause where it looks at book_ride as

Cue 227 — 00:16:23,544 --> 00:16:26,486

destructive, which means that it can cost money.

Cue 228 — 00:16:26,526 --> 00:16:30,886

So confirm again. Yeah. This is the interrupt with that I was highlighting earlier

Cue 229 — 00:16:30,926 --> 00:16:35,086

for a tool call where we are marking it as destructive.

Cue 230 — 00:16:35,126 --> 00:16:40,726

You can give it something called interrupt width where it knows that if it's trying

Cue 231 — 00:16:40,766 --> 00:16:45,926

to do a destructive action, it interrupts its own flow and responds back to the user

Cue 232 — 00:16:45,966 --> 00:16:50,386

asking, hey, do you want to confirm if this is what you actually want?

Cue 233 — 00:16:51,006 --> 00:16:54,946

Again, this is something we were not able to do previously.

Cue 234 — 00:17:00,226 --> 00:17:01,426

I can talk about this a little bit.

Cue 235 — 00:17:01,466 --> 00:17:07,226

So in the previous case, we didn't have the ability to store history of these

Cue 236 — 00:17:07,266 --> 00:17:10,906

conversations. It was just a bunch of API logs.

Cue 237 — 00:17:10,946 --> 00:17:15,346

But now the entire conversation itself is the state.

Cue 238 — 00:17:15,386 --> 00:17:19,926

So you don't have to separately maintain a history somewhere else.

Cue 239 — 00:17:21,026 --> 00:17:22,426

The AI pretty much takes care of it.

Cue 240 — 00:17:22,466 --> 00:17:27,666

And this is the one I was talking about with tool interrupts where if you call

Cue 241 — 00:17:27,706 --> 00:17:34,666

something interrupt with allows you to pause the loop as it's working and ask

Cue 242 — 00:17:34,706 --> 00:17:39,066

the user to whether to confirm or deny the request, and then continues on in the

Cue 243 — 00:17:39,106 --> 00:17:41,926

same path without breaking the conversation chain.

Cue 244 — 00:17:43,106 --> 00:17:47,586

Now, where if if everything is so good.

Cue 245 — 00:17:47,626 --> 00:17:49,946

What are some of the things we're giving up?

Cue 246 — 00:17:49,986 --> 00:17:53,586

So the main thing of giving up is determinism, right?

Cue 247 — 00:17:53,626 --> 00:17:57,866

Previously, because it was so hard coded and because it was the mapping was so

Cue 248 — 00:17:57,906 --> 00:18:02,346

rigid, you knew exactly which phrase calls which tool.

Cue 249 — 00:18:02,386 --> 00:18:09,146

Whereas now it's it's completely left on the agent's interpretation of

Cue 250 — 00:18:09,186 --> 00:18:14,466

the English language itself. Now you can do certain things to improve that on your

Cue 251 — 00:18:14,506 --> 00:18:19,706

side, which we have seen increased efficiency for like having better guardrails or

Cue 252 — 00:18:19,746 --> 00:18:25,946

eval checks. You can have it make sure that it doesn't end up in a constant loop of

Cue 253 — 00:18:25,986 --> 00:18:29,286

guessing itself with max turns.

Cue 254 — 00:18:30,466 --> 00:18:37,386

Understanding when to insert the interrupt with a tool to make

Cue 255 — 00:18:37,426 --> 00:18:41,246

sure that it's not able to do something that it's not supposed to do.

Cue 256 — 00:18:41,946 --> 00:18:48,546

There is also latency. Like, it depends on how well the agent can understand,

Cue 257 — 00:18:48,586 --> 00:18:52,826

like the from the point of you asking the agent a request to the point of delivering

Cue 258 — 00:18:52,866 --> 00:18:57,666

on your request. That is entirely dependent on how well it understands the context

Cue 259 — 00:18:57,706 --> 00:19:00,826

in the first place, and how well your guardrail and evals are designed.

Cue 260 — 00:19:00,866 --> 00:19:06,266

Whereas previously it was very deterministic, you could you could measure it out and

Cue 261 — 00:19:06,306 --> 00:19:09,926

improve the performance of it pretty deterministically.

Cue 262 — 00:19:14,286 --> 00:19:20,106

Okay. So essentially where this is going is.

Cue 263 — 00:19:21,846 --> 00:19:26,286

We have been migrating a lot of things, even in in the work that I'm doing at my

Cue 264 — 00:19:26,326 --> 00:19:33,206

company, where we are using now instead of hard-coded API calls in terms of user

Cue 265 — 00:19:33,246 --> 00:19:37,406

requests, we are making tool calls for most of those.

Cue 266 — 00:19:37,446 --> 00:19:41,926

So my company is currently doing HR tech platform stuff.

Cue 267 — 00:19:41,966 --> 00:19:46,726

So let's say I want to hire somebody or I want to send somebody an offer letter.

Cue 268 — 00:19:46,766 --> 00:19:51,466

I can ask my agent in Bill to like, hey, send this person an offer letter.

Cue 269 — 00:19:51,686 --> 00:19:58,246

It knows which calls to make to grab that data and then what tools to

Cue 270 — 00:19:58,286 --> 00:20:01,426

execute to actually send that information out.

Cue 271 — 00:20:04,166 --> 00:20:05,986

Do you want to add anything to it?

Cue 272 — 00:20:09,946 --> 00:20:13,306

I mean, that's basically the the MCP route.

Cue 273 — 00:20:13,346 --> 00:20:19,866

So previously columns are tools, but now we basically have some sort of an MCP layer

Cue 274 — 00:20:19,906 --> 00:20:23,924

that allows agents to understand which tools are available to them.

Cue 275 — 00:20:23,964 --> 00:20:25,924

Yeah. Yeah, that's all we had.

Cue 276 — 00:20:25,964 --> 00:20:28,824

But we have time for questions.

Cue 277 — 00:20:29,044 --> 00:20:32,604

We're happy to answer a question on, like, this new agent world,

Cue 278 — 00:20:32,644 --> 00:20:35,544

how different agents are talking with each other.

Cue 279 — 00:20:36,324 --> 00:20:37,784

MCP, etc..

skills

talk-sharma-mohapatra-evolving-intent

tile.json