MAIO

LangGraph from LangChain
Amazon Bedrock’s AI Agent framework
Rivet (drag and drop GUI)
Vellum (GUI for complex workflows)
n8n (drag and drop GUI)

Frameworks simplify standard tasks but add abstraction layers
Abstractions can obscure prompts/responses and make debugging harder
Can encourage unnecessary complexit

• Basic building block: LLM enhanced with augmentations
  • Retrieval
  • Tools
  • Memory
• Current models can actively use these capabilities
  • Generate search queries
  • Select appropriate tools
  • Determine what information to retain
• Implementation focus:
  • Tailor capabilities to specific use cases
  • Ensure easy, well-documented interface

• Decomposes tasks into sequential steps
• Each LLM call processes output from previous one
• Can add programmatic checks between step

When to use: When tasks can be cleanly decomposed into fixed subtasks

Example use cases:

• Generating marketing copy then translating it
• Writing document outlines then creating full documents

Retrieval
Tools
Memory

Generate search queries
Select appropriate tools
Determine what information to retain

Tailor capabilities to specific use cases
Ensure easy, well-documented interface

Decomposes tasks into sequential steps
Each LLM call processes output from previous one
Can add programmatic checks between steps

When tasks can be cleanly decomposed into fixed subtasks

Example use cases:
Generating marketing copy then translating it
Writing document outlines then creating full documents

Classifies input and directs to specialized followup tasks
Allows for separation of concerns and specialized prompts

When to use: For complex tasks with distinct categories better handled separately
Example use cases:
Directing different customer service query types
Routing simple questions to smaller models, complex ones to larger
models

Workflow: Parallelization
Two key variations:
Sectioning: Breaking tasks into independent parallel subtasks
Voting: Running the same task multiple times for diverse outputs

When to use: When subtasks can be parallelized for speed or when multiple perspectives improve confidence

Example use cases:
Sectioning:

• Implementing separate guardrails and core responses
• Automating evals for different aspects of performance

Voting:

• Code vulnerability reviews from multiple prompts
• Content moderation with different evaluation thresholds

Central LLM dynamically breaks down tasks
Delegates to worker LLMs
Synthesizes results

When to use: For complex tasks with unpredictable subtasks
Example use cases:
Coding products requiring changes to multiple files
Search tasks gathering information from multiple sources

One LLM generates responses
Another provides evaluation and feedback in a loop

When to use: With clear evaluation criteria where iterative refinement adds value
Example use cases:
Literary translation with nuanced requirements
Complex search tasks requiring multiple rounds of refinement

Understanding complex inputs
Reasoning and planning
Using tools reliably
Recovering from errors

1. Initial command or discussion with user
2. Independent planning and operation
3. Gaining ”ground truth” from environment at each step
4. Pausing for human feedback when needed
5. Task completion or reaching stopping conditions

Best for:
Open-ended problems with unpredictable steps
Situations where hardcoding paths isn’t possible
Tasks requiring multiple LLM operational turns
Trusted environments where scaling is important

Higher costs
Potential for compounding errors
Need for extensive testing in sandboxed environments
Appropriate guardrails

Customer Support:

• Natural conversation flow with external information access
• Tool integration for customer data and knowledge base
• Programmatic actions (refunds, ticket updates)
• Clear success measurement through resolutions

Coding Agents:

• Verifiable solutions through automated tests
• Iteration based on test feedback
• Well-defined and structured problem space
• Objectively measurable output qualit

Think about the model’s perspective when designing tools
Include example usage, edge cases, and format requirements
Name parameters clearly with precise descriptions
Test extensively and iterate based on model usage
”Poka-yoke” your tools to make mistakes harder

Give the model enough tokens to ”think”
Use formats similar to natural internet text
Avoid formatting overhead like line counting or string escaping

1 Maintain simplicity in your agent’s design
2 Prioritize transparency by showing planning steps
3 Carefully craft your agent-computer interface (ACI)
Remember:
Success isn’t about the most sophisticated system
It’s about the right system for your needs
Start simple and add complexity only when it demonstrably improves outcomes

Competitive strategy is concerned with ”creating and maintaining a competitive advantage in each and every area of business

Competitive strategy concerns the firm‘s actions that make it outperform its competitors

Competitive Advantage Derives from Cost Leadership and Differentiation

A company prices around the market average, and enjoys superior profits because its costs are lower than those of its rivals

A company adds value in areas of significance for the customer, who then accepts a premium price for distinctiveness of products and services.

Size differences and economies of scal

Experience and learning economies

Low-cost access to factors of production

Technology advantage

Policy choices

The cost advantages are obtained due to size, or scale of operation, with cost per unit of output decreasing as fixed costs are spread out over more units.
→ Apply to manufacturing, sales, marketing, & administration

- Ability to purchase specialized and costly technology
- Division of labor and specialization
- Shared overhead cost