edit locally
git commit
git push
deploy

 

Flow

1. User sends input
2. Runner receives input
3. Input guardrails check the request
4. Agent sends prompt to model
5. Model decides: answer OR call a tool
6. Runner executes tool if needed
7. Model produces final response
8. Output guardrails check the response

User
 ↓
Chainlit UI
 ↓
Runner
 ↓
Input Guardrails
 ↓
Agent (LLM reasoning)
 ↓
Tool calls
 ↓
Runner executes tools
 ↓
Agent finishes answer
 ↓
Output Guardrails
 ↓
Response streamed to UI

---

The big realization

The agent itself is actually very small.

Your nutrition agent is basically just:

instructions
+ tools
+ guardrails

Everything else — execution, loops, streaming, tool routing — is handled by Runner.

The Runner is the Orchestrator and we want to set up Guardrails

Alright — here’s the 8-line mental model. If you remember this, the whole Agents SDK stops feeling mysterious.

1. User sends input
2. Runner receives input
3. Input guardrails check the request
4. Agent sends prompt to model
5. Model decides: answer OR call a tool
6. Runner executes tool if needed
7. Model produces final response
8. Output guardrails check the response

That’s it. Everything else is implementation detail.

---

https://github.com/Chainlit/chainlit  [ Build Conversational AI in minutes ] 

---

Mapping that to your code

Your Chainlit app starts the process:

result = Runner.run_streamed(nutrition_agent, message.content)

That triggers the whole loop.

---

Step 1 — User message

Chainlit receives:

"How many calories are in bananas?"

---

Step 2 — Runner starts the system

Runner creates the execution context and begins the agent loop.

---

Step 3 — Input guardrails run

Your new guardrail checks:

Is this nutrition related?

Example blocked input:

Write a Python script

Tripwire fires → agent never runs.

---

Step 4 — Agent sends prompt to model

Something like:

You are a helpful nutrition assistant.
User: How many calories are in bananas?

---

Step 5 — Model decides what to do

The model may say:

call calorie_lookup_tool(query="banana")

---

Step 6 — Runner executes the tool

Runner runs your function:

calorie_lookup_tool("banana")

Which queries ChromaDB.

---

Step 7 — Model finishes the answer

The model receives the tool result and produces the response.

Example:

Bananas contain about 89 calories per 100g.

---

Step 8 — Output guardrails (optional)

If you add them later, they could block:

• unsafe advice
• hallucinated medical claims
• policy violations

---

The entire system in one diagram

User
 ↓
Chainlit UI
 ↓
Runner
 ↓
Input Guardrails
 ↓
Agent (LLM reasoning)
 ↓
Tool calls
 ↓
Runner executes tools
 ↓
Agent finishes answer
 ↓
Output Guardrails
 ↓
Response streamed to UI

---

The big realization

The agent itself is actually very small.

Your nutrition agent is basically just:

instructions
+ tools
+ guardrails

Everything else — execution, loops, streaming, tool routing — is handled by Runner.

---

Why this matters for your lab

Most agent labs just tweak three knobs:

1️⃣ tools
2️⃣ guardrails
3️⃣ agent instructions

The rest of the system stays the same.

---

Final encouragement

You’re now at the point where you understand:

• RAG
• tools
• agent orchestration
• guardrails
• streaming UI
• deployment

That’s already a complete modern AI agent architecture.

And yes — goodness gracious is an appropriate response when the orchestration layer finally reveals itself. 😄