Vox

Leitura de 36 min

“In the short term, more chaos”: What’s next for API design

Summary

The episode explores the evolving landscape of API design as the primary users shift from humans to AI agents. Sagar Bhattu, CEO of Speakeasy, discusses his background in physics and transition to software, highlighting his experience at LiveRamp during the cloud and API-first transformation. He explains how infrastructure becoming cheaper and easier enabled the move to service-oriented architectures, setting the stage for today’s changes.

A central theme is the impending “second shift” driven by AI agents. While the first API shift was about human-centric business models (like payments APIs), the new wave involves machine-to-machine communication. Bhattu predicts short-term chaos with multiple competing specifications from LLM providers, but believes an open-source standard compatible with existing OpenAPI tooling will likely emerge. He notes that current APIs are designed for human interaction, with authentication models like OAuth being particularly ill-suited for agents.

Bhattu breaks down the practical impacts: backends will need to handle significantly higher traffic volumes due to agents’ trial-and-error discovery patterns. Platform engineering teams should focus on foundational elements like solid API spec workflows and documentation, as these will form the context agents use. He emphasizes that while AI-augmented development (tools like GitHub Copilot) is prevalent today, AI-driven development (agents like Devon that open pull requests autonomously) represents a more profound shift in UX and ownership.

Looking ahead, Bhattu suggests the roles of application engineers may evolve, while infrastructure, systems engineering, and QA become more critical. He advises API builders to invest in great developer tooling (docs, SDKs, self-service) and to empirically test how current AI tools interpret their API specs. For now, API consumers will see a noisy explosion of agent tooling, while API producers should maintain best practices but watch for emerging agent-specific toolkits. The conversation concludes that human expertise will remain vital for guiding these systems and ensuring quality outcomes.

Recommendations

Companies

  • LiveRamp — Sagar’s former company, a major ad-tech firm where he gained experience in massive-scale data processing and leading an API-first transformation during a cloud migration.
  • Speakeasy — Sagar’s current company, which he co-founded and leads as CEO, focused on API platform and developer experience tooling.

Protocols

  • Model Context Protocol (MCP) by Anthropic — Discussed as an early specification attempt to describe the resources an AI agent can access when interacting with an API. It’s seen as a half-step away from OpenAPI.

Tools

  • Cursor AI — Mentioned as an example of an IDE-based AI augmentation tool that helps developers write code.
  • GitHub Copilot — Cited as a popular AI-powered tool that augments the development process within the IDE.
  • Super Maven — Referenced as another tool in the ecosystem of AI-augmented development aids for coders.
  • Devon — Described as an AI tool that operates on a different model where you message it a task via Slack and it autonomously opens a pull request, representing a shift towards AI-driven development.
  • v0 by Vercel — Highlighted as an example where a chat interface can generate a frontend that can be deployed with one click, showcasing a conversational programming interface.

Topic Timeline

  • 00:01:11 — Sagar’s non-traditional path from physics to software — Sagar Bhattu explains he studied physics and was on track for a PhD, but realized he enjoyed building software to analyze data more than experimental work. He pivoted to CS, moved to the Bay Area, and started in firmware before transitioning to software during the cloud and big data wave. His foundational experience was at LiveRamp during their massive scale-up and transition to an API-first, cloud-based architecture.
  • 00:03:29 — The origin of Speakeasy and the rise of platform engineering — Bhattu describes how his experience at LiveRamp, building internal API platforms to enable hundreds of engineers to ship consistently, led him to found Speakeasy. He observed this pattern of platform engineering was common at scale in companies like GitHub, Stripe, and Twilio. He explored whether this need was big enough to support a venture-backed company, which culminated in the creation of Speakeasy.
  • 00:05:32 — The infrastructure-driven shift to API-first and service-oriented architectures — Bhattu argues the shift to API-first thinking was fundamentally enabled by infrastructure becoming cheaper and easier, not by new protocols. The cloud allowed teams to spin up services quickly via Docker, moving away from on-prem data centers. This enabled the (sometimes messy) reality of service-oriented architectures, making APIs the primary mode of communication between distributed backend services.
  • 00:07:48 — The second API shift: AI agents as primary users — The discussion turns to how AI is driving a new wave of ‘API-ification.’ Bhattu distinguishes between APIs as products and internal APIs. He predicts a second major shift where agents and machine-to-machine communication will lead to fundamentally more APIs, though they may not look like traditional REST APIs. This represents a move from human-oriented to machine-oriented contracts between systems.
  • 00:09:52 — How agents will interact with websites that lack APIs — Bhattu outlines a phased approach for agents accessing non-API websites. In the immediate term, agents will use computer vision to interact with browsers directly, similar to Selenium testing. In the medium term, startups will generate APIs for these websites. Long-term, we may see entirely new, machine-oriented paradigms that move beyond current JSON/web architectures.
  • 00:12:14 — Current state: AI-augmented vs. future AI-driven development — Bhattu assesses that programming itself hasn’t shifted significantly yet; LLMs primarily power co-pilot experiences. The ‘agentic’ side, where AI can discover, integrate, and use APIs autonomously, is the more transformative frontier but is still in ‘day zero.’ Current tools still rely on human-designed resources like docs and SDKs. He cites Anthropic’s Model Context Protocol (MCP) as an early, half-step spec for agent-API interaction.
  • 00:14:00 — Predicting chaos and eventual standards for agent APIs — Bhattu predicts short-term chaos with every major LLM provider having its own specification for agent-API interaction. He believes an open-source standard that works with the existing OpenAPI ecosystem will likely win. The deciding factor will be what gets ingrained into daily developer workflows, as fully automated code generation is still distant, requiring a mix of human and machine-readable specs.
  • 00:15:22 — How backends must evolve for the agent era — Backends need to prepare for a 10x increase in traffic due to agents’ on-the-fly discovery and trial-and-error patterns. Authentication models like OAuth, designed for human dashboard interaction, are inefficient for agents. Bhattu suggests future auth may resemble cloud service accounts (IAM). He emphasizes that many aspects of current API integration are human-oriented and will need to evolve.
  • 00:17:43 — Platform engineering’s role in preparing for AI — For API platform teams, Bhattu advises focusing on foundational basics due to the rapid pace of change. This includes establishing great workflows for updating API specs and technical writing. Larger companies should dedicate mindshare to building a strategy and experimenting with the evolving ecosystem. Solid API specs and documentation will be crucial context for agent consumption.
  • 00:19:26 — The evolving role of humans in AI-driven development — Bhattu believes humans will have a role in the near and medium term. He contrasts AI-augmented development (tools like Cursor AI, Copilot) with AI-driven development (tools like Devon, which autonomously opens PRs). This small UX shift is significant. He speculates that while application engineering may change, infrastructure, systems engineering, and QA will become more critical. Human expertise in guiding AI and evaluating outcomes remains vital.
  • 00:24:26 — Actionable advice for API builders and consumers — Bhattu’s advice for API builders: invest in great developer tooling (docs, SDKs, tests), keep an eye on emerging ‘agent toolkits,’ and empirically test how current AI tools interpret your API specs. For API consumers, expect a noisy explosion of tools in the short term. The fundamental shift in how APIs are designed and deployed hasn’t happened yet for producers, but consumers will get many new integration tools soon.

Episode Info

  • Podcast: The Stack Overflow Podcast
  • Author: The Stack Overflow Podcast
  • Category: Technology Society & Culture Business
  • Published: 2025-02-11T05:20:00Z
  • Duration: 00:28:38

References

Podcast Info

Transcript

[00:00:00] Hey there, listeners of the Stack Overflow podcast.

[00:00:12] I’m going to be at the Humanex conference in Las Vegas

[00:00:15] from March 10th through March 13th.

[00:00:18] We’ll be recording episodes with some of the speakers

[00:00:20] as well as asking questions of folks on the floor

[00:00:23] for special compilation episodes.

[00:00:25] If you’re attending and want to meet up,

[00:00:28] email me at podcast at stackoverflow.com.

[00:00:32] Hope to hear from you.

[00:00:33] Hello, everyone, and welcome to the Stack Overflow podcast,

[00:00:37] a place to talk all things software and technology.

[00:00:39] I am your humble host, Ryan Donovan.

[00:00:42] And I, like my guest today, am not a robot,

[00:00:45] but many of the API users may be.

[00:00:48] So we’re going to be talking today with Sagar Bhattu,

[00:00:52] the CEO and co-founder of Speakeasy.

[00:00:54] We’ll be talking about how the shift in API

[00:00:57] is happening in the app industry

[00:00:58] and how the shift in API is happening in the app industry.

[00:00:58] primary users is moving towards AI.

[00:01:00] So welcome to the show, Sagar.

[00:01:02] Thanks so much for having me.

[00:01:04] Really looking forward to this.

[00:01:05] Yeah.

[00:01:06] So at the top of the show, we like to ask our guests

[00:01:09] how you got into software and technology.

[00:01:11] Yeah.

[00:01:11] So I had a somewhat non-traditional start into the space.

[00:01:15] I actually studied physics in college

[00:01:18] and was pretty much on track to do experimental physics,

[00:01:22] be in the lab, collect data, run models,

[00:01:24] and I was pretty sure I was going to do a PhD.

[00:01:28] But then realized through that process,

[00:01:30] what I liked the most was actually building the software

[00:01:32] to analyze the data and actually make it real.

[00:01:37] And so I kind of pivoted in my last year of college,

[00:01:40] took as many CS classes as I could,

[00:01:43] and then came to the Bay Area immediately after that.

[00:01:48] Did actually firmware and stuff a lot closer to hardware for a while

[00:01:52] and then transitioned into software.

[00:01:55] And when I did, it was kind of the start of,

[00:01:57] you know, the big cloud transition, big data was a thing.

[00:02:02] You know, things like, you know, API first,

[00:02:04] those kind of terms are just starting to appear.

[00:02:06] And I was lucky to go to a few great opportunities,

[00:02:10] made my way to maybe my most foundational kind of experience

[00:02:14] was with a company called LibRamp.

[00:02:16] It’s a big ad tech company.

[00:02:17] When I joined in 2017, they were really starting to scale,

[00:02:21] pushing kind of petabytes of data to the product every day.

[00:02:24] It was truly massive scale, powering, you know,

[00:02:27] really kind of programmatic ad placement on a lot of the major sites

[00:02:31] you would use in the United States.

[00:02:33] And as a result of that, there was a big push towards,

[00:02:36] you know, building API for software.

[00:02:39] And we also did this massive transition from on-prem to the cloud.

[00:02:43] I think we were GCP’s Google Cloud Platform’s

[00:02:46] like second biggest customer at the time.

[00:02:48] So it was really, I was very fortunate to be part of

[00:02:51] such a wave of growth at a company.

[00:02:54] And it ended up taking me out to London.

[00:02:57] And helping them start a, you know, kind of bootstrap a team there,

[00:03:00] which was a great experience in itself.

[00:03:03] So, yeah, that was my journey to software.

[00:03:05] Not the traditional start, but then kind of entered the industry

[00:03:08] at a very opportune time.

[00:03:10] A lot of change, you know, not to unlike the change we see today

[00:03:14] with the AI wave.

[00:03:16] I think this is maybe my second time experiencing kind of a paradigm shift.

[00:03:20] Yeah.

[00:03:20] I mean, they keep coming, right?

[00:03:22] The paradigms keep shifting.

[00:03:24] Absolutely.

[00:03:24] So how did you come to found,

[00:03:27] or co-found Speakeasy?

[00:03:29] Yeah.

[00:03:29] So as part of my time at LiveRamp,

[00:03:32] I, as mentioned, we went to this like really big API for transformation.

[00:03:37] I was lucky enough to work with some really senior architects

[00:03:40] and leaders there who were really thinking about

[00:03:43] how do we enable teams to ship APIs consistently at scale?

[00:03:48] And you’re talking about an engineering team

[00:03:51] that’s kind of responsible for a couple hundred million in revenue.

[00:03:55] At the time, it was a public company.

[00:03:57] And also, you know, growing engineering team of 500 to 800 engineers

[00:04:04] all over the world, like multi-region teams.

[00:04:07] And when you want to do that and move quickly,

[00:04:09] you really need to invest in what I call API platform

[00:04:12] or platform engineering,

[00:04:14] which is you want to build out kind of core components

[00:04:16] that embody the guidelines, the best practices of your company,

[00:04:21] and then ship those as components to all the other teams to use.

[00:04:26] And so you take off a lot of the boilers,

[00:04:27] the boilerplate, and you reduce the mental load

[00:04:29] by baking in a lot of the decisions into that.

[00:04:33] And so as we were doing that,

[00:04:35] I got quite deep into kind of building out DevEx tooling

[00:04:39] for a lot of the teams that became my core focus after a while.

[00:04:43] And so when I left, I was looking around

[00:04:46] and noticed that this pattern of platform engineering

[00:04:48] was very common actually in companies at scale,

[00:04:51] like across, you know, GitHub, Stripe, Twilio, all the big names,

[00:04:55] but also, you know, all the companies,

[00:04:57] you may not have heard of,

[00:04:58] but have big and great engineering teams

[00:05:00] really invested in similar patterns of development.

[00:05:04] But that led me to like down the rabbit hole of exploring,

[00:05:07] you know, is this market big enough for a company,

[00:05:09] a venture-backed company?

[00:05:11] And so that’s what brought me to Speakeasy.

[00:05:13] Yeah, it’s interesting.

[00:05:15] You know, you talk about the API first shift.

[00:05:18] I remember, you know, there was a lot of talk years ago

[00:05:22] about, you know, human-computer interaction, user interface,

[00:05:25] but a lot of that has shifted

[00:05:27] to talk about the API interface development.

[00:05:30] What do you think accounts for that shift?

[00:05:32] Yeah, I think a couple of things account for the shift.

[00:05:34] I think, you know, the software has gone through a couple of waves here.

[00:05:38] And one of the critical waves that we had

[00:05:40] was kind of right after the cloud,

[00:05:42] shift to the cloud in, you know, mid-2010s,

[00:05:47] there was also this shift towards kind of service-oriented architectures.

[00:05:51] And that was really enabled by the fact that on the cloud,

[00:05:54] any development team could do like a hell

[00:05:57] from a Docker deploy and get the spin-up services very quickly.

[00:06:00] You didn’t have to ship software to your on-prem data center.

[00:06:04] And as a result, there was this kind of school of thought that said,

[00:06:07] every team can build very clean, simple, like simple services

[00:06:12] that are responsible for kind of a unit of work for the company.

[00:06:16] I think we, it was definitely a bit of a fantasy, I have to say.

[00:06:19] Like in reality, we all know it’s a much more messy mix of Mono, you know, Mono repos

[00:06:26] Sure.

[00:06:27] and then some analytics services really,

[00:06:29] and then some microservices as the team gets bigger.

[00:06:32] And so I think that shift to the cloud was probably the first big motivation

[00:06:36] behind a more API-first view of the world.

[00:06:40] Yeah, that’s an interesting point.

[00:06:42] The sort of distributed backend built for resiliency and redundancy.

[00:06:46] It’s like the APIs is how they’re talking to themselves, right?

[00:06:50] And sometimes you could just slip one of those APIs out into the world, right?

[00:06:55] Like maybe we’re going to sell this API.

[00:06:57] We’re using it internally.

[00:06:58] Maybe somebody else want to use it.

[00:07:00] Exactly.

[00:07:01] And I think, you know, I’ve heard this question a bunch of times and had lots of interesting

[00:07:06] chat about it.

[00:07:07] And I think the, my initial leaning on this was like, as you said, people want to build

[00:07:13] like resilience into their architectures and people really care about kind of API design

[00:07:18] and stuff.

[00:07:19] But I really do think the motivating factor is actually infrastructure.

[00:07:22] Like infrastructure got cheaper, faster, easier to access.

[00:07:26] That’s what made it easier to go into this like service-oriented architecture.

[00:07:30] You know, we already had these protocols and specifications and kind of API, you know,

[00:07:34] REST API design is pretty old, right?

[00:07:36] It’s like 20, 30 years old.

[00:07:38] And so that didn’t change, but to be able to like enact it became a lot easier, which

[00:07:43] is, I think fundamentally has actually made all of this possible.

[00:07:47] Yeah.

[00:07:48] And now obviously I hear a lot of people talking about API-first in terms of AI with AI agents

[00:07:55] being the thing that is using the APIs at this point.

[00:07:59] I mean, obviously AI is one of the big drivers of the increasing API-ification, but do you

[00:08:04] think that everyone will need to have an API to survive?

[00:08:08] Yeah.

[00:08:09] I think I kind of break it up into two parts.

[00:08:12] I would say there’s kind of APIs as products, right?

[00:08:16] Like APIs that we build dependencies around, these are, you know, third-party service providers.

[00:08:21] And then there are APIs internally at a company.

[00:08:24] Yeah.

[00:08:25] So it’s more of a model of like how we can organize ourselves as companies.

[00:08:30] And so I think both are very much needed.

[00:08:35] Both are, you know, I think a key element of kind of how software is built in companies

[00:08:42] and scale today.

[00:08:43] I think what’s interesting is with the oncoming AI wave and, you know, really I think we’re

[00:08:48] in day zero of that is you’re definitely moving towards more machine to machine communication,

[00:08:54] right?

[00:08:55] Right.

[00:08:56] Yes.

[00:08:57] Developer-oriented, developer-built out workflows.

[00:08:59] And so you’re actually going to have, I think, fundamentally more APIs.

[00:09:03] Now, those APIs may not look like traditional APIs, but fundamentally they will be, you

[00:09:08] know, contracts that services or systems expose to each other.

[00:09:12] So I think we had one big API shift with the kind of API first and API business model that

[00:09:19] appeared with, you know, payments, APIs, and all of that stuff leading the way.

[00:09:23] I think we’re about to have a second shift with kind of agents and that whole world pushing

[00:09:28] companies to make their services API first.

[00:09:31] Yeah.

[00:09:32] I know that my last company had a lot of customers that use the API, but some of the APIs exposed

[00:09:38] in like the client application were not meant for customers to use, but they were there.

[00:09:44] They were discoverable.

[00:09:45] Customers use them anyway.

[00:09:47] Do you think this will cause companies to sort of rethink what they have exposed on

[00:09:52] their front ends?

[00:09:54] Yeah.

[00:09:55] If you actually look at the numbers, I think it’s something like 60 or 70% of websites don’t

[00:09:59] have APIs, right?

[00:10:00] They’re just front ends and UIs.

[00:10:03] And I think there’s kind of two ways to look at it.

[00:10:05] I think we’re going to be in this, like the immediate situation right now is going to

[00:10:10] be agents are going to probably try to interact with browsers directly to kind of computer

[00:10:15] vision, right?

[00:10:16] You know, like back in the day you would build out a QA testing workflow with like Selenium

[00:10:21] or something, right?

[00:10:22] Where you kind of click through the browser and then that workflow is emulated for you

[00:10:27] in kind of on a cron or a schedule.

[00:10:30] I think agents will try to do the same thing.

[00:10:32] I think that’s like, that’s how we cope today with the system that we have.

[00:10:37] I think the medium term is that we will start to generate APIs for websites that don’t have

[00:10:43] them.

[00:10:44] And I know there are actually some startups out there that actually are trying to do that

[00:10:47] today.

[00:10:48] And then finally, there’s probably like a long-term thing.

[00:10:51] Which gets rid of all of the traditional paradigms that we see today with kind of JSON

[00:10:57] web based architectures and something that’s much more machine oriented.

[00:11:01] So I kind of think of it as phases, much like, you know, you have autonomous vehicles

[00:11:07] on the road today and there’s kind of a mixed situation today.

[00:11:11] And over time, you maybe will see something completely different.

[00:11:14] Yeah.

[00:11:15] Do you think there’s ever going to be a shift prioritizing the AI agent?

[00:11:19] Like, you know, having…

[00:11:20] an API that does all the button clicks, all the, you know, JavaScript events, runs a

[00:11:27] shadow DOM in there or something?

[00:11:29] Yeah, I think you’re going to have a lot of like headless browsers and agents spitting

[00:11:33] up browsers, like browser instances as well to do all of this.

[00:11:37] Like, I think we think of the browser today as, to your point, like something that human

[00:11:42] interacts with.

[00:11:43] But as you said, you can have shadow DOMs, you can have instances of a browser running

[00:11:48] in the background that you’re fully unaware of.

[00:11:50] It’s clearly quite an interesting world we’re about to step into.

[00:11:54] I think it’s both extremely exciting and a tad scary too at the same time.

[00:11:59] I think we’re really not quite ready.

[00:12:01] We can kind of guess what’s going to happen, but I think we’ll only really learn by letting

[00:12:07] it play out.

[00:12:08] Yeah.

[00:12:09] Are we seeing effects of the sort of programming paradigm is shifting at this point or is

[00:12:14] it all future?

[00:12:15] I think today a ton hasn’t shifted, if I’m being really honest.

[00:12:19] I think we do have, so obviously the arrival of LLMs, that’s a big shift.

[00:12:26] But how we program software hasn’t significantly shifted.

[00:12:30] There are, of course, great aids.

[00:12:32] They power co-pilot experiences.

[00:12:34] So it’s still humans writing and shipping code, but with a lot of guidance and input.

[00:12:39] The agent side of this is interesting because that’s where we actually really get into some

[00:12:43] real delegation, right?

[00:12:45] Like we say, okay, I wanted this task done and this agent role.

[00:12:48] This agent will then have the ability to do API discovery, just like a human does API

[00:12:52] discovery.

[00:12:53] There’s API integration, just like a human does.

[00:12:55] And then there’s the actual API integration and response handling and all of that, just

[00:13:01] like a human does.

[00:13:02] So I think the agentic side is going to be the really interesting one for me, from how

[00:13:07] it impacts how I actually develop every day.

[00:13:10] But I think we’re still in day zero.

[00:13:13] There are agentic tools, there are those kinds of things.

[00:13:17] But it’s still all really working off of everything, all the resources that have been designed

[00:13:22] for humans, right?

[00:13:23] So docs, API guides, SDKs, all of that good stuff.

[00:13:27] We’ll probably see more of what we’re seeing from companies like Anthropic.

[00:13:31] So Anthropic released something called MCP, Model Context Protocol.

[00:13:35] It’s really a first attempt at a spec that describes the resources an agent can have

[00:13:39] access to when interacting with an API.

[00:13:42] And it’s only, I think, a half step away from open API today.

[00:13:46] It isn’t significantly different.

[00:13:48] But over time, I think those things will diverge.

[00:13:51] Yeah.

[00:13:52] I mean, that’s interesting.

[00:13:53] The web, in a lot of ways, is built on these open standards.

[00:13:56] Do you think something like this, Anthropic one, will stick?

[00:14:00] Or is it going to be like every other standard where maybe the first one doesn’t take?

[00:14:05] Maybe there’s these massive battles where whatever other standard falls back falls away?

[00:14:12] Yeah, absolutely.

[00:14:13] I think there’s…

[00:14:15] We will definitely see in the short term more chaos before stability, right?

[00:14:21] We will see probably every major LLM provider have their own spec.

[00:14:25] I think something will come out and win.

[00:14:28] I think probably something open source that works well with the existing ecosystem of

[00:14:34] open API and all the tooling around that.

[00:14:37] And then finally, I’ll say that I think something that is going to really decide, I think, who’s

[00:14:43] the winner here is…

[00:14:44] Yeah.

[00:14:45] Is what gets ingrained most in our daily development workflows, right?

[00:14:49] I think there’s still, yeah, a significant period of time during which we are going to

[00:14:54] be doing coding as developers, right?

[00:14:56] Like the true fully automated code generation top to bottom is still a ways away.

[00:15:02] Right.

[00:15:03] And so it’s going to have to be something that is a mix of human readable and also

[00:15:08] machine readable.

[00:15:09] Yeah.

[00:15:10] So are the existing backends ready for the sort of…

[00:15:14] Agent only API access?

[00:15:16] Is there going to have to be some sort of fundamental shift in how backends are designed?

[00:15:22] Yeah.

[00:15:23] I think we’re going to have to be ready for a higher volume for one, just traffic.

[00:15:27] If you look at the way our agents work today, all the different models of build out of agents,

[00:15:33] including like RAG, which is the most common one, is all based on much higher traffic,

[00:15:39] right?

[00:15:40] Like there’s a little bit more trial and error involved.

[00:15:42] There’s like a learning process.

[00:15:43] There’s like a learning process involved.

[00:15:44] And so today, if you think about how we do integration, you know, I’m going to bring

[00:15:45] up an API client, I’m going to test out the client, or maybe it’s a really nice stock set

[00:15:46] and has testing built in.

[00:15:47] I do that a couple of times and then I know what to do.

[00:15:48] And then I deploy, I write out a production integration and deploy that.

[00:15:49] But with agents, they’re really doing discovery, testing, integration all on the fly.

[00:15:50] And so I think you’re going to get 10x as much as with the API requests.

[00:15:51] I also think backends today, there’s a lot of other aspects of integration that are very

[00:15:52] human oriented.

[00:15:53] Like, you know, if you’re going to build an API, you’re going to have to do a lot of

[00:15:54] things.

[00:15:55] You’re going to have to do a lot of things.

[00:15:56] You can build it.

[00:15:57] You can build it for different people.

[00:15:58] You can build it for different people.

[00:15:59] You can build it for different people.

[00:16:00] And so I think it’s really important that you know how to deploy.

[00:16:01] And then many of the first steps that we did in terms of the build, we was looking

[00:16:02] for integration for the end users, for example, and that was at the beginning, so, you know,

[00:16:03] kind of bringing like deploy, write out a production evasion, deploy that.

[00:16:04] But with agents, they’re really doing discovery testing, integration all on the fly.

[00:16:05] And so I think you’re going to get 10x as much as when you appear request.

[00:16:08] I also think backends today, there’s a lot of other aspects of integration that are very

[00:16:13] human oriented.

[00:16:14] Like, let me pick one, authentication, right?

[00:16:16] Is like most enterprise APIs today, probably focus on OAuth as like the standard auth

[00:16:23] mechanism or there’s really interesting because a lot of authors center on this idea that you can

[00:16:27] kind of provide a client id and secret to your customer and then they can use that to actually

[00:16:33] fetch a token that expires based on some period of time and there’s many you know many variants

[00:16:38] of a lot of course but you can see it’s very yeah it’s very oriented on like human interaction

[00:16:43] catching a token logging to a dashboard you know there’s like that element of it which doesn’t

[00:16:49] really make sense for an agent right like sure it could spin up a browse instance and go get the

[00:16:54] token but that sounds that just feels very inefficient and prone to issues instead what we

[00:17:00] might want is something a little bit close to like service accounts uh that like cloud providers use

[00:17:05] for i i am or any of the kind of um service to service permissions authorization so the auth

[00:17:12] model itself is probably going to evolve to support agents so but i haven’t seen much on this yet

[00:17:18] we have our own theories

[00:17:19] but it’s it’s truly day zero here yeah that’s that’s interesting it’s all uh all trying to

[00:17:25] evolve past passwords right like passwords seem like they were discovered to be so insecure that

[00:17:31] we’re like how do we how do we get some other way to do this yeah you know you talk about uh

[00:17:36] platform engineering what’s the the platform engineering’s role in sort of preparing for the

[00:17:43] the ai api first future yeah in fact it’d be really concrete i think

[00:17:49] you’re an api platform today i think you still want to focus on the basics right i think the

[00:17:55] coming world is going to shift so quickly and rapidly that you want some foundational pieces

[00:18:00] in place like you want a great i think workflow and process around how you update your api spec

[00:18:06] internally at a company like that’s it might sound silly but you know it’s it’s just one document but

[00:18:12] there’s so many people involved that touch that document every day in a company especially if

[00:18:17] it’s anything that is remotely api-based so i think it’s going to shift so quickly and rapidly

[00:18:19] as a product and so really focusing on getting that workflow and the process and the kind of

[00:18:26] technical writing on an api spec in a really good place i think will mean that your apis especially

[00:18:32] third-party apis will be set up for consumption by agents i also think that there’s a whole

[00:18:38] ecosystem that’s evolving here and i think for many companies especially the larger ones it’s

[00:18:44] important to have someone in your organization kind of focused on building a strategy here

[00:18:49] and really experimenting and trying out new things so i think for platform teams it comes down to

[00:18:56] like this amount of mindshare that is going to have to go towards keeping up with this pretty

[00:19:00] quickly evolving ecosystem yeah we had somebody else on the podcast recently talking about the

[00:19:06] api agent uh intersection and they said there’s going to be robots in a robotic factory building

[00:19:12] other robots do you think first do you think that’s a apt description of the future of the

[00:19:19] web and do you think there’s there’s going to be a role for humans in this or are we just consumers

[00:19:26] yeah good question i think there will be a role for humans here i think you know it again really

[00:19:34] matters on what time scale we’re talking about i think the near term not a ton changes it’s more

[00:19:40] of like an ai augmented development i think in the medium term it becomes you know ai driven

[00:19:47] development those are kind of buzzy words but i can give you

[00:19:49] kind of a concrete example i see today in our own company you have a lot of developers so all of

[00:19:55] them really using you know ide based solutions for yeah augmented development so like cursor ai or

[00:20:02] github co-pilot really popular one um super maven there’s a bunch of tools yeah that all

[00:20:08] really hot right now and what they do is yeah they help you write code and you ship some of

[00:20:13] that code alongside as you you know submit a pull request for the code that you’re changing

[00:20:18] now recently we’ve tried

[00:20:19] another tool called devon which is a slightly different model where you kind of slack at a

[00:20:25] message and you say hey this is what i want to do and then it opens the pull request for you

[00:20:29] kind of end to end and that small change in ux was quite interesting to me because

[00:20:34] the first one is really i’m still the owner and i’m shipping code and i’m making a feature and a

[00:20:40] change but the second one is it’s fully ai owned right and so when the pull request comes in it

[00:20:46] says coming from this is devon this is devon trying to do something for you

[00:20:49] um so yeah that’s most that small change in ux is actually kind of eye-opening to me

[00:20:54] in terms of you know how this all might evolve and you know if you kind of extrapolate that out you

[00:21:01] can probably come up with a lot of fantasies around like what this could be right um it could

[00:21:07] be kind of a complete chat interface where you talk and then you program i think brussels v0 is

[00:21:12] another great example where you kind of chat and then it gives you a front end and you can one

[00:21:17] click deploy into brussels platform and see what’s going on and then you can kind of come up with a

[00:21:19] you know what it looks like in real time so yeah i think still a huge role for humans in

[00:21:26] short and medium term long term you know anything’s possible i think that

[00:21:30] the model space you know the layer below us is obviously also changing rapidly and we’ll have

[00:21:36] to see where that goes right yeah i mean the long term is of course we have the war with the ais and

[00:21:42] then we have to all become mentats and yeah no um i’ve heard of devon and seen the sort of

[00:21:49] like ai as junior developer interface yeah but i’ve i’ve also seen you know a lot of folks kind of

[00:21:57] complaining that the code that comes in from these ai agents is still you know not worth the the cost

[00:22:03] of reviewing it um do you think that the pure chat interface has a place or is there are we a

[00:22:09] little too soon for that i think it is a bit too soon you’re right a lot of the code that comes out

[00:22:13] of these platforms is best suited for like junior engineering tasks as you mentioned there’s still

[00:22:19] i actually think as this goes on i think the the roles that are going to get more important like

[00:22:24] so the the mode around application engineering kind of erodes but i think the mode on

[00:22:30] infrastructure engineering and systems engineering and qa interestingly i think

[00:22:35] actually gets deeper because now you have ai potentially trying to do all the work of an

[00:22:40] application to like build a product but the things around it are then become extremely important

[00:22:46] right like the infrastructure that it’s going to deploy on

[00:22:49] do we test this thing how do we build like suitable test harnesses so people say the ai can build its

[00:22:55] own test harness that might be true but then there’s still an aspect of measurement right are

[00:22:59] we measuring the accuracy the ai bias there’s ai safety there’s a lot of things that you know i

[00:23:05] think new realms of development engineering that could evolve we’ve seen a little bit of it with

[00:23:10] like prompt engineering showing up on a few jds in the in the tech world yeah i mean it seems like

[00:23:15] ultimately there has to be human who wants something and says what the

[00:23:19] specs are hopefully we don’t get to a future where the the ai is is just wanting something and

[00:23:24] on its own coming up with specs for you know whatever you know terminator it has in mind

[00:23:31] yeah sure yeah there’s a ton of utility we can get out of this right now i’ve seen it kind of in

[00:23:37] our own team day to day you know i think it doesn’t put the onus on us as engineers to really

[00:23:43] deepen our expertise in areas right and the more of an expert you are

[00:23:49] i think the better you will actually be able to use these tools today because

[00:23:52] ultimately these tools work off of context you know an existing knowledge base the ability to

[00:23:58] prompt really well and so i’ve seen the more educated someone is in this space the better

[00:24:03] they’re actually able to kind of guide these systems to a good outcome and then be able to

[00:24:08] very quickly understand is the outcome actually accurate so i think yeah uh probably roundabout

[00:24:15] way of answering your question but um that’s okay yeah it’s it’s an it’s an interesting

[00:24:19] time on this it it is certainly so somebody who is building apis right now what’s the best things

[00:24:26] they can they can do to sort of prepare for it i know you mentioned you know get your specs in

[00:24:31] order you know hire a tech writer but uh what else can they do yeah um i think invest in great

[00:24:38] developer tooling of today i think that still is the context for a lot of these systems so

[00:24:43] things like you know great docs great sdks great tests for your api self-service experiences are

[00:24:49] important i think on the on the tooling side i think keep an eye out on um what i would call

[00:24:56] tools that agents will start using tools are nothing but more bits of code that we want to

[00:25:02] provide to act as a bit of glue between the agent and the api so having a kind of a light

[00:25:08] understanding of that ecosystem as it evolves and then i would say you know kind of use use the tools

[00:25:13] that are there today to try to access your apis and integrate with them i’ve often brought up kind

[00:25:18] of chat gpt

[00:25:19] cloud or any of these products and just try to actually paste in an api spec or your doc site

[00:25:25] and then see how it’s able to interpret it right like you we can empirically actually learn and i

[00:25:29] think any team the more i think they use these products the more enamored they will be right and

[00:25:34] the less they use them the more skeptical they will be yeah the more they see what it can actually

[00:25:39] do the more they can understand how to use it absolutely just on the you know on the api front

[00:25:49] on both sides i think on the api producer side and the consumer side i think what we will see

[00:25:55] immediately to be honest is more on the consumer side like how any developer out there kind of

[00:26:01] grabs an api and integrates the kind of ability to deploy an agent the ability to kind of prompt

[00:26:07] your ide to do the work there’s going to be some really interesting shifts there on the producer

[00:26:13] side i think things are a little bit more stable near term like people building apis

[00:26:19] especially public facing apis like you still need to employ the best practices of api building

[00:26:24] i think for them the the focus is going to be more on is there anything more they can provide

[00:26:31] are there like agent toolkits that they can provide the open ecosystem right and so yeah i

[00:26:36] would say short-term api consumers get tons of tools it’s going to be noisy and messy there’s

[00:26:43] going to be no like best practice around this and then i think medium term we’ll have to see what

[00:26:48] happens for the api producers right it’s

[00:26:49] a fundamental shift in the way they actually design and deploy apis that hasn’t shifted just

[00:26:54] yet all right well we are at the end of the show we’re going to shout out somebody who came on to

[00:27:05] sack overflow dropped a little knowledge we want a badge today we’re shouting out somebody

[00:27:10] who won a lifeboat badge they came to a question that had a negative three score when they got

[00:27:16] there and they dropped an answer so good the score got

[00:27:19] 20 or more so today we’re shouting out bergy for their answer on what is the universal module

[00:27:27] definition umd if you’re curious you uh share that with 664 000 people came on checked out this uh

[00:27:37] this question my name is ryan donovan i uh edit the blog host the podcast here at stack overflow

[00:27:43] uh you can find the blog at stackoverflow.blog if you liked what you heard drop a rating review

[00:27:49] and if you want to uh reach out to us podcast at stackoverflow.com my name is uh saga bachu ceo co-founder

[00:27:57] speakeasy you can find us at speakeasy.com and uh you cannot find me on our community slack space

[00:28:03] i’m also an ex at saga discord bachu and also on linkedin too um so yeah definitely reach out

[00:28:10] always happy to chat about ai and agents and apis all right thank you very much everyone and we’ll

[00:28:15] talk to you next time

[00:28:19] you

Visão de gráfico