APIs for AI: Have we failed?

2024 Nordics APIs Platform Summit Keynote

https://www.youtube.com/watch?v=lsdizCmFs7U
https://nordicapis.com/sessions/apis-for-ai-have-we-failed/

Transcript

1. Zdeněk “Z” Němec, superface.ai, October ’24 APIs for AI: Have we failed? superface.ai
2. Shall we stop building APIs? I am here to tell you that we should probably stop building APIs unless we change how we do it. superface.ai
3. Zdeněk “Z” Němec • designed an API description language • pioneered API-design & API-first • built APIs with the largest enterprises • building API infrastructure for AI agents • building AI agents Founder and CTO of Superface.ai • DSL director Apiary.io → Oracle • Founder of Good API consulting • Enterprise API strategy and execution superface.ai
4. LLM-powered Agents superface.ai
5. LLM-powered Agents • An agent is a software that acts on behalf of user or another program • LLM-powered agents can independently • perform tasks • make decisions • interact with other systems • Agent needs tools to be able to do meaningful things superface.ai
6. Tools LLM-powered agents • Agents are only as good as the tools they have • computation • databases • API calls • LLM works more like a human • It’s able to fi gure out what endpoints to call from API documentation • Agents are more resilient to API changes LLM-powered Agent Memory Planning Tools APIs superface.ai
7. You would think the APIs are the best things for agents superface.ai
8. In many situations it is easier, faster and cheaper to access data and perform actions via UI over API. Often, it is the only way. superface.ai
9. LinkedIn A minute vs. several weeks • LinkedIn API exists • It is not easily accessible to developers • The functionality is limited • Using 3rd party scraping service • Connect my agent to LinkedIn in one minute superface.ai
10. Instagram Two minutes vs. several weeks • API exists but it is not accessible • Some of our most popular articles are regarding getting the access to API • Using Apify web actor we are able to read data from Instagram in two minutes superface.ai
11. Accessing a Salesforce instance of a biggest HW maker Deal or no deal • Entering data into CRM • The case was saving 38hrs / week • Agent did not get the API access to the CRM • Entering the data using web actor (scraper) works superface.ai
12. There are many more similar stories • No access to API for agent developers • Legacy APIs (WS/RPC) • API can do only fraction of what UI can • Overcomplicated APIs superface.ai
13. Jesse Lyu, Founder and CEO @rabbit “We don’t like to work with APIs because you are betting everyone will give you an API which is not the case. It’s easier for OpenAI to get an API built or open for them, it’s hard for startups to convince the API providers.” https://www.youtube.com/watch?v=X-MNgciL5hw superface.ai
14. Problems with APIs and AI APIs are slowing down AI agents • API does not exist • You won’t get access • The functionality is not there • It’s impossible to convince companies to built functionality for you • Companies have no incentive to do it • → API Economy superface.ai
15. API Economy Business drivers for building APIs Innovation & e ffi Value add Partners Stickiness Direct monetization superface.ai
16. API Economy Business drivers for building APIs Innovation ffi Value add Partners Stickiness Direct monetization the case for AI superface.ai
17. API Economy Business drivers for building APIs Innovation ffi Value add Partners Stickiness Direct monetization what are agents challenging: commoditizing your SaaS apps superface.ai
18. API Economy Business drivers for building APIs Innovation ffi Value add Partners Stickiness Direct monetization what are agents challenging: commoditizing your SaaS apps B2B agents
19. API Economy Business drivers for building APIs Innovation ffi Value add Partners Stickiness Direct what are agents challenging: commoditizing your SaaS apps B2B agents agents purchasing capabilities
20. API Economy Business drivers for building APIs Innovation ffi Value Partners Stickiness Direct what are agents challenging: commoditizing your SaaS apps B2B agents agents purchasing capabilities agents adding value on top of API
21. Attention economy AI challenging the stickiness • Salesforce launched its "End of Software" campaign in 2000 • 25 years since the inception of SaaS / Cloud • Most users don’t like to work with CRM (and software ) • No.1 use-case for Superface agent is dealing with CRM • But companies want us to stick in front of the screen, with their UI superface.ai
22. “Governance” problem AI challenging the value add • Limited data integration • Power plant operators have to observe the dashboards and react when something happens • A “scraping” solution was developed because of the lack of data integration • A camera observing the dashboard with image recognition alerts the workers when a reaction is needed superface.ai
23. There is one more problem with AI and APIs… superface.ai
24. LLM performs miserably finishing the complex tasks with complex APIs superface.ai
25. Agents are the future of API consumption superface.ai
26. The way agents have to interact and do business with APIs is broken superface.ai
27. If you can’t use APIs… superface.ai
28. LAM: Large action models superface.ai
29. • Instead of tools, LAMs use web and screen actors to interact with systems on your behalf • Similar to robotic process automation, but • more capable (reasoning, semantics inference) • in fi nitely more reliable (no hard-coded UI) • usable by end-users (users can add actions) • LAM interprets user commands and execute tasks LAM: Large Action Model Neuro-symbolic engine with actors superface.ai
30. Rabbit.tech How LAM works • Record users performing actions e.g. “using Uber Eats to order meatballs” • Record enough people over time • LAM understands and knows what am I doing, where am I clicking • Then, the action is invoked using voice via R1 device • Rabbit’s neuro-symbolic LAMs executes the action https://www.rabbit.tech/research
31. Open Interpreter “LAM” approach • Runs directly on your computer • Controls your computer and applications • Can be used with voice device communicator • Uses your identity since it runs on your behalf • You can “teach” it your work fl ows using voice https://github.com/OpenInterpreter/open-interpreter superface.ai
32. • Companies are scared of loosing attention • APIs do not exist • APIs are limited • APIs have a poor design LAMs running the actions reliably, without the need for APIs -vs- this wasn’t feasible before LLMs superface.ai
33. “Engineering of APIs is an unnecessary complication if all the data and actions can be achieved using the UI.” superface.ai
34. superface.ai We could be doing so much more useful work with agents but we are not because of the obstacles.
35. Have we failed with APIs? superface.ai
36. We need APIs more than ever! superface.ai
37. Get APIs right or people will either use LAM or walk away. ( and agents ) superface.ai
38. APIs are built for the old world • Agents are your new best friends • As API provider / product owner go out and look • The API consumption will be 80% agents and 20% humans • Do everything to support LLM-powered agents • API- fi rst • API-design for AI • Streamline API access • Drive the business incentives superface.ai
39. superface.ai The way humans and software are interacting has changed.
40. Attention economy will shift. • my agent does not care about your commercial • is my agent your daily active user? • what is your KPI? superface.ai
41. We won’t be manually integrating your API. We won’t be sitting in front of your UI. superface.ai
42. Thank you! Let’s talk Zdenek “Z” Nemec 
 Twitter/X: @zdne LinkedIn: https://www.linkedin.com/in/zdne/ superface.ai superface.ai
44. LLM function calling And API design • Mechanism that LLM uses to recognize a and call a function (tool) • Ideally, we would give an API to LLM and it would fi gure it out • The more atomic and “normalized” the API is the harder it is to use • Complex tasks and atomic (high granularity) APIs makes it di ffi cult for LLMs LLM detects function requirements Structured function input Execution of the function Returning the result API call LLM function calling repeat multiple times superface.ai
45. UI vs. API APIs and Web • Hypertext (HTML) is a representation of resources • Originally a resource could have multiple representations (e.g. HTML and JSON) • APIs parted away from the original REST API concept • We are stuck with Web and APIs providing di ff erent capabilities • What is available via API is now a product (and business) decision Contact resource superface.ai
46. CAPTCHA and other anti-scraping • LLM is actually helping me to beat the anti-robot protection • anti-scraping race is something you don’t have to do if you have a solid API program superface.ai
47. https://www.reddit.com/r/Rabbitr1/comments/1fu5gst/ is_lam_still_a_thing_true_innovation_or_a_lie/ “Go to target.com and order 2 boxes of Fancy Feast canned cat food for me" 1. Opens the webpage 2. Searches for products 3. Adds them to the cart 4. Enters the order interface At the fi nal checkout interface, my input is required to con fi rm the order
48. Practically speaking How to minimize the need for LAM • API- fi rst • Provide the same data and actions as UI* • API product management • API design for AI: bundle atomic calls for complex task • Streamline access delegation • Make APIs access accessible • Security team should be helping secure the access not ensuring they are covered superface.ai