fill

AGENTIC DEVELOPMENT IN PRACTICE

ABOUT ME

  • Experimentator by heart.
  • Keeping my personal home laboratory.
  • Building the TORQ AI project.

WHY AGENTIC DEVELOPMENT MATTERS NOW

  • AI: Models can act, inspect, and iterate autonomously.
  • Scaling: Systems become too complex for human oversight.
  • Velocity: Fast and reliable dev cycles.
  • Quality: Enforcement of patterns and standards.

TOOLING

Each tool offers different approaches to AI-assisted development with varying levels of autonomy and integration.

  • CodeMie
  • Cursor
  • GitHub Copilot

PET PROJECT IDEA

ABOUT STRAVA

  • You train (running, riding, or anything else).
  • The activity becomes public.
  • You get more motivation to train!

ABOUT TORQ

  • Turns Strava activities into art.
  • No manual prompts anymore!
  • Makes better visual feedback from the Strava tape.
diagram

LET'S TRY IT!

https://torq.club

AI 101

CONTEXT = MEMORY

  • Input context: Everything the model can see and reference.
  • Working memory: Current conversation and file contents.
  • Limitations: Token limits and attention mechanisms.
  • Strategy: Carefully curate what context gets included.

SYSTEM PROMPT = POLICY LAYER

  • Behaviors: How the agent should operate.
  • Guardrails: What the agent should and shouldn't do.
  • Personality: Tone, style, and approach preferences.

PROMPT = RUNTIME CONFIGURATION

  • Instructions: How you want the model to behave.
  • Task: What exactly needs to be accomplished.
  • Constraints: Boundaries and requirements for the output.
  • Examples: Templates and patterns.

TOOL = FUNCTION CALLING

  • Actions: Predefined functions the model can invoke.
  • Effects: File operations, API calls, system commands...
  • Results: Tool call outputs.

MODELS ARE STATELESS

  • No persistence: Each interaction starts fresh.
  • Context-dependent: Behavior is based on the context.
  • Reproducible: Same inputs produce similar outputs.
  • Orchestration: External systems manage state and continuity.

AGENTIC DEVELOPMENT 101

fill

CODE EDITOR INLINE SUGGESTIONS

  • Autocomplete++: Context-aware code completion.
  • Single-line focus: Predict what you're about to type.
  • Limited scope: Works within current file.
fill

FILE-SCOPED REASONING

  • File-level: Comprehend file structure and purpose.
  • Context-aware: Consider relationships between code blocks and files.
  • Interactive: Back-and-forth conversation.
fill

ORCHESTRATED AGENT

  • Structured: AGENTS.md.
  • Skills: Composable behaviors.
  • Hooks: Automated responses.
  • Systematic: Consistent patterns.

AGENTIC DEVELOPMENT SCHOOLS

fill

VIBE-CODING

  • Intuitive: Based on feel.
  • No planning: Agent guides the architecture.
  • Rapid: Fast feedback loops.
  • Trade-offs:
    • Speed vs. structure.
    • Flexibility vs. consistency.
fill

SPEC-DRIVEN DEVELOPMENT

  • Doc: Specification before implementation.
  • Contracts: Upfront behaviors.
  • Systematic: Structured planning.
  • Predictable: Reduced surprises.

SKILLS & HOOKS

fill

SKILLS: REUSABLE CAPABILITIES

  • Standard: Predictable actions for common tasks.
  • APIs: Act like function calls for the agent.
  • Knowledge: Capture best practices in reusable form.
fill

SKILLS.SH

The easiest way to start with skills is to use shared ones!

fill

HOOKS: EVENT-DRIVEN TRIGGERS

  • Automated: React to system actions.
  • Feedback: Continuous monitoring.
  • Guardrails: Prevent common mistakes early.
fill

AI Templates

The easiest way to start with hooks is to use shared ones!

fill

VIBE CODING TIME!

PRACTICAL THINGS

  • Plan with heavy model, implement with simple one: Use powerful models (GPT-4, Claude) for architecture and planning, lighter models for implementation
  • Divide and conquer: Break complex tasks into smaller, focused contexts with clear boundaries

THE END

The real shift is not AI writing code.
It's engineers designing systems that include AI as an actor.

fill

QUESTIONS TIME!