1
Question intake
Student asks a mixed FAQ and account question.
All samples · Agent orchestration
Multi-step Service Triage Agent
A tool-calling pattern that chooses FAQ lookup or human escalation instead of answering everything directly.
Problem
During the hackathon, students ask repeated schedule and submission questions, but account or secret issues need tutor escalation.
Users
Student teams, tutors, and operations staff handling D-day support.
Why this track
This is the in-curriculum chance to practice ReAct-style orchestration and tool boundaries without building a large agent platform.
Architecture
Stay minimal. 5-6 nodes. Each arrow is one network hop.
user
Student question
agent
gpt-4o-mini agent loop
tools
Tool schema registry
faq
lookup_faq()
human
escalate_to_human()
state
Team issue state
Edges
- user agent — question + context
- agent tools — tool choice
- tools faq — public topic
- tools human — private or blocked issue
- faq agent — tool result
- human state — ticket note
Prompt Pack
Starting prompts. Iterate. Move the system prompt into prompts/system.md so it can be versioned.
system
You are a hackathon service triage agent. Use lookup_faq for public schedule, submission, and cost questions. Use escalate_to_human for account, key, private data, or unresolved issues. Never claim you fixed an account yourself. user
Our app cannot call Azure and we also need to know when submission closes.Code Snippet
The pattern shape. Read it, run the matching scaffold, then adapt the idea for your own team.
python
first = client.chat.completions.create(
model=deployment,
messages=messages,
tools=TOOLS,
)
call = first.choices[0].message.tool_calls[0]
result = run_local_tool(call.function.name, call.function.arguments)
messages += [first.choices[0].message, tool_message(call.id, result)]
final = client.chat.completions.create(model=deployment, messages=messages)
# ... your turn: add one safe tool for your domain
Reference: src/techniques/agent_orchestration/ in halla-ai/hackathon-sample-2026
Demo Screens
Three screens that prove the prototype works.
2
Tool decision
Debug panel shows FAQ lookup plus tutor escalation note.
3
Student reply
Final answer separates public guidance from tutor follow-up.
Azure budget
Uses gpt-4o-mini chat plus tool calls. Keep to one or two tool turns per request; most demos stay below USD 1.
Pitfalls
- • Symptom: agent loops. Cause: no maximum tool-call count. Fix: cap calls and return a tutor handoff.
- • Symptom: wrong tool arguments. Cause: vague schema. Fix: use short enums and required fields.
- • Symptom: model reveals hidden state. Cause: streaming internal steps. Fix: show only user-safe tool summaries.
Possible Extensions
If you finish the 1-day path early, use one question below to make the project more original.
- How would your version remember unresolved issues across team members?
- How would tutors approve or reject an escalation before it becomes visible?
- How would you prevent the model from selecting a tool for a private action?