The End of
User Centric
Product Development

Software agent, a computer program that performs various actions continuously and autonomously on behalf of an individual or an organization.

A Generative AI Agent
is defined as an application 
that tries to achieve a goal by
observing the world and acting

upon it using the tools it

has at his disposal​

Ich möchte nächstes Wochenende von Berlin nach Hamburg fahren. Suche mir einen Bus bei Flixbus heraus.

DemoTime

Conversational Agents

  • Chat with your documents
  • Q&A, Knowledge Interactions
  • User queries (like NQL) :-)
  • fulfil user transactions
     
  • Everything CoPilot, ChatGPT, ...
     
  • Chat, Voice, Mail. 

Chatbots are the Ford Model T of AI

  • This is not just a faster horse.
  • It is a lot more flexible and configurable
  • It is faster, can carry heavier weights, more comfortable
     
  • The actual revolution was still something else: automation

Autonomous /
Workflow Agents

  • Limited User Interaction
  • Event driven, Queues of Tasks
  • Chains of task execution

Human in the loop

Ask and remember

How to build
an Agent

Perception & Tools

Right now: 

  • all documents
  • all images
  • all sounds
  • all apis
  • all databases
  • all services 

Soon:

  • your desktop
  • your phone
  • robots

Strategy & Planning

  • o1-preview: evaluate plans before acting
     
  • Who taught o1 the strategies?
     
  • Yep, we did.
     
  • We can do this for
    agents, too.

Decisions

  • Did we reach our goal?
     
  • What tool should we try next?
     
  • Are we in a loop?
     
  • Is the result of the tool helpful?

Memory

  • Short Term Memory
    • "What do i need to know
       for this decision"
    • "What do i need to
       plan my next steps"
    • Mostly prompt
       
  • Long Term Memory
    • Which strategy actually worked out?
    • which strategy went bad?
    • Which steps were useless?
    • Mostly vectordb

Reflection and Learning

The reason why Strawberry's R are not a problem anymore

 

  • Reflection                                                               

  • Feedback loops

  • Backtesting

  • RLHF

Flow Engineering

In Agentic AI,
user activity
is waste
expensive.

User Story Mapping?
Impact Mapping

How to capture requirements
without any users ...

Todo:
The Cognitive Architecture
 

  • Strategy/Planning

  • Decisions and Knowledge

  • Reflection and Learning

  • Self-Optimization / Generalisation        

"As an Agent, i want
to remember working
 strategies to learn."

We haven't solved this one yet.

BUT ...

You never know if you can rely on AI ...

Backtesting, Learning.

"Create a product vision for the company replacing flixbus."

"Decide which is the most important app. Provide UX and user journey concept."

 

"Create a prototype for this user journey."

React-Prototype after 30 seconds:

This could have
been 100 Click-
Dummies

It's not human,
some tasks need empathy.

It's just a fad.
It won't stay.

“The horse is here to stay, but the automobile is only a novelty – a fad.”


President of the Michigan Savings Bank, advising Horace Rackham (Henry Ford’s lawyer) not to invest in the Ford Motor Company, 1903

  • Automatisierung der Produktion:
    10.000 Fahrzeuge pro Tag
     
  • Ein Ford Model T wurde in
    93 Minuten vollständig produziert
     
  • Die Preise konnten in 5 Jahren von
    950$ auf 360$ gesenkt werden

Industrialisierung

The industrialization of
knowledge work:
elastic knowledge work

  • 100 times more expensive than software
  • 100 times less expensive than humans.
     
  • If you got one agent, you got 1000, too.
  • Failing 20 times before it works is still cheap.

Helpful metaphors

  • 1000 smart & eager to work interns:

    What should they do for you?

     
  • If you had 100 hours instead of one
    for the next task, what would you do?

Horace Rackham still invested
5.000$ in Ford stock and sold them
later  for 12.5000.000$


 

You could use a car instead of a horse, too.
 

Or even scale it and travel by bus or train.

Automating Knowledge

Work