Building Your Own Agentic Framework

by 🎙️ @aakash7539 & @stalwartcoder
  • 🧑🏻‍💻 Software Engineer at Tune AI
  • 🏗️ Building GenAI, ML Systems, and Infra at Tune
  • @aakash7539
    
    

About Aakash

About Abhishek

  • 🧑🏻‍💻 DevRel Engineer at Tune AI
  • 🏗️ Been building backend systems
  • @stalwartcoder

Why Agents?

  • Solve Complex, Multi-step tasks
     

  •  Proactive vs. Reactive
    (Agents vs. ChatGPT)
     

  • 24/7 Operation vs. Single-interaction model

 

*Email*
follow this thread

*Analyst*
Hey, i found the
discrepancies

  • "Just a software" that uses A.I. to accomplish tasks.

  • Executes multiple steps.

  • Steps are determined by an LLM.

  • Sometimes, the “orchestration layer” is needed.

What ❓

“A person's a person, no matter how small.”

Dr. Seuss

Agentic Workflow

  • Sophisticated iterative and multi-step process to interact and instruct LLMs to complete complex tasks with more/better accuracy.
  • In this process, a single task is divided into multiple small tasks that are more manageable
     
  • And each task leave scope for improvements throughout the task completion process.

Multi agent framework in a nutshell

 

  • Proactive Not Reactive

  • Non-stop (Internal monologue)

  • Commodities Tools

 

What Tune thinks Agents should be

Internet Search 🔍

  • SerpAPi
  • Exa.ai
  • anon.com

Maps 🗺️📍

  • Google Maps
  • OSM

Documents 📃

 

  • Google Docs
  • Notion

Key Components 🏗️⚙️

  • Function-calling LLM 🧠

  • Task schedulers and orchestrators 🤹🏻‍♂️

  • Tool integration for expanded capabilities 🛠️

Types of Workflows

Agent Server Example

Limitations

  • Models are "stupiddd"
    • Recursive multistep sucks
    • Hallucination with scaling functions
    • Baby level understanding of complex functions - data problem
    • Massive models
       
  • LLMs are not fast (Not enough tokens/sec)
    • ​Agents        Thinking
    • Thinking       Tokens
    • Less Tokens       Less thinking/sec
    • Less thinking/sec       Slow agents

Limitations

  • Personalised Intelligence
    • TV vs YouTube
    • Your values are not reflected in the agents you use
       
  • Tools for humans not AI
    • Tools are Not API first

Text

Learnings

  • Help your model - "only" add what you need
  • Keep function names simple
    • func attendAConferenceForMe ❌
    • def attend_conf ✅
  • Be mindful about your function arguments
  • Simpler data types
    • structs, classes  ❌
    • int, float, string ✅
  • No sensitive functions
    • delete_user
  • Split your complex tasks

Checkout the more samples 👇

Subscribe to our newsletter for the latest in GenAI news and updates 📰 🤓

 tune.beehiiv.com

Thank You!!

deck

By Abhishek Mishra