AI Architecture
Agentic AI & Multi-turn Interactions
Moving beyond simple "prompt-response" models to autonomous, goal-oriented systems that maintain state
and iterate.
1. Agentic AI
Agentic AI refers to AI systems that act as agents—they don't just answer questions;
they perform actions to achieve a goal. They possess:
Core Characteristics
- Autonomy: Ability to make decisions without constant human intervention.
- Tool Use: Capability to use external tools (APIs, databases, code interpreters)
to gather information or affect the world.
- Planning: Breaking down complex goals into a sequence of executable steps
(Chain of Thought).
- Reflection: Reviewing outputs and self-correcting errors.
- Memory: Maintaining context over long periods (Short-term & Long-term).
Goal: "Deploy this app to Cloud Run"
|
v
[Agent Core]
|--> Plan: 1. Build Docker image -> 2. Push to GCR -> 3. Deploy
|--> Action: Run `docker build` (Tool: Shell)
| |
| (Error: Missing Dockerfile)
| |
|--> Reflection: "I need to create a Dockerfile first."
|--> Action: Create Dockerfile (Tool: FileSystem)
|--> Action: Run `docker build` again
|--> Action: Run `gcloud run deploy`
v
Result: Success
2. Multi-turn Interactions
Multi-turn capability allows an AI to maintain a coherent conversation over multiple
exchanges, remembering previous context, refining answers, and handling complex workflows that cannot be
solved in a single shot.
Comparison
Single-turn (Stateless)
User: "Fix this code."
AI: "Here is the fixed code."
End of logic.
Multi-turn (Stateful)
User: "Fix this code."
AI: "I see a bug. Fixing it..."
User: "It's still failing test X."
AI: "Ah, I missed that edge case. I recall you mentioned X earlier. Updating..."
3. Applied to Nostra (Our Context)
How we leverage these concepts in the AI-Boosted Market Creation feature:
Workflow
- User Intent: "Create a market for the election."
- Agent Action: "I need more details. Which election? What candidates?"
(Proactive questioning).
- User Response: "2026 Seoul Mayor."
- Agent Action: Uses
Search Tool to find candidates (Oh Se-hoon,
etc.).
- Agent Action: Drafts market parameters (Title, Outcomes, Description).
- User Feedback: "Add another candidate." (Multi-turn refinement).
- Agent Action: Updates draft and executes
Create Market Tool.
← Back to Index