← Reference documents
End-to-end runbook

Agentic Gathering Runbook (End to End)

The full playbook for a Sponic Agentic Gathering: from hardware prep and member interviews through the live dinner event to post-dinner testimonials. Covers what Team, Agent, and Members do at every stage.

Runbook Phase 1 · Tech test Source: meeting 2026-05-08
Phase 1 goal: This is a technology test. The priority is discovering what works and what doesn't — not polish. Expect things to break; that's the point. Fix forward.
0 / 0 tasks
Contents
  1. Overview & three archetypes
  2. Phase A — Prep (days before)
    1. Hardware setup
    2. Software setup
  3. Phase B — Member interviews (day before / morning of)
  4. Phase C — Day of the event (chronological schedule)
  5. Agent prompts (toggle)
  6. Phase D — Post-dinner
    1. Testimonials & feedback
  7. Related docs

Overview & three archetypes

Every person at the Agentic Gathering interacts with three layers of the experience. Throughout this doc, tasks are tagged by who owns them:

ArchetypeWhoRole summary
Team Sonia + engineering Setup hardware, configure software, run dry-runs, operate devices during event, collect feedback, compile results.
Agent AI systems (interview agent, Agentic Gathering manager, host agent, survey agent) Conduct interviews, parse profiles, generate playlist & cook guide, host the Agentic Gathering conversation, run feedback survey, compile event summary.
Member Dinner guests / Sponic members Complete interview, wear mic, participate in dinner, give feedback after.

Phase A — Prep (days before the event)

1. Hardware setup

Each attendee needs their own device + microphone channel. No shared mics — one person, one device, one audio stream for clean speaker isolation.

ItemQuantityNotes
Mac laptops (host + backup)2–3One Mac is the Agentic Gathering host, mirrored to TV. Others for team ops.
Phones (Android / iOS)1 per guestRun the Sponic app for subtitles / translation and audio input. Get builds from the builds page.
Lavalier mics — Saramonic set1 setClip-on, wired to phone. Good isolation.
Lavalier mics — DJI set1 setWireless. Test range + latency before event.
AirPods / BT earbuds2–3 pairsFallback for guests without a dedicated mic. Test pairing day-of.
TV / external display1Mirror the host Mac so all guests see the AI host's output.
Speaker (optional)1For Spotify playlist playback. Bluetooth or AUX from host Mac.
Chargers + power stripsPlentyLiveKit burns ~10–15% battery per 30 min. Keep devices plugged in.
Mic strategy for Phase 1: Test multiple mic types (AirPods, Saramonic, DJI, built-in phone mic) across different people to learn what works best. Don't optimize yet — gather data.

Team tasks

2. Software setup

Pre-event dry runs

Phase B — Member interviews (day before / morning of)

Each guest is interviewed by the AI interview agent before the Agentic Gathering. The interview collects personal context that feeds the Agentic Gathering experience: food preferences, interests, conversation style, languages spoken.

Flow

  1. Member Opens the Sponic app on their device, taps "Interview" (or equivalent).
  2. Agent The interview agent greets them and asks questions guided by the interview prompt.
  3. Member Talks naturally. 10 min max. One speaker, one mic.
  4. Agent Interview ends. Raw transcript is saved to voice_notes.
  5. Agent Transcript parser runs:
    • Profile fields — maps answers to structured fields in app_users (dietary restrictions, languages, interests, etc.).
    • Summary text — a distilled summary of interview points (not the raw transcript). This is what the Agentic Gathering LLMs will see.
  6. Agent All context used to build the profile is stored and visible to the member — full transparency.
Design decision: LLMs receive the processed summary, not the raw transcript. People repeat themselves, go on tangents — the summary gives the agent clean, structured context.

Trigger: Agentic Gathering manager

Once all (or most) interviews are complete, the Agentic Gathering manager agent activates. It can be triggered by a webhook on interview completion or run as a recurring job (e.g. hourly). It:

  1. Collects all member summaries.
  2. Generates a Spotify playlist tailored to the group's tastes.
  3. Generates a cook dietary guide — what each person can/can't eat, allergies, restrictions. Sent to the cook with enough lead time to shop and prep (ideally day before).
OutputWhen readyDelivered to
Cook dietary guideDay before (or several hours before)Cook / host
Spotify playlist~1 hour before dinnerHost Mac / speaker

Phase C — Day of the event (chronological schedule)

Below is the minute-by-minute schedule from when guests arrive. Each time slot shows what Team, Agent, and Member should be doing.

T − 60 min — Setup
Venue & tech prep
Team
Team
Team
Team
Agent Dinner party manager has already generated the playlist + cook guide. No active agent tasks.
T − 30 min — Final checks
Last-mile tech validation
Team
Team
Team
Team
Agent Standing by. Host agent prompt loaded and ready.
T − 0 — Guests arrive
Welcome & onboarding
Team
Team
Team
Member
T + 10 min — Dinner party begins
AI host goes live
Team
Team
Agent Dinner party host agent activates. Reads all member profiles/summaries. Opens with a warm greeting. Begins steering conversation.
Member Sit at the table. Mic on. Look at the TV for the host's output. Speak naturally. Use personal device for subtitles/translation.
T + 10 to T + 40 min — Main dinner conversation
Live AI-hosted dinner (~30 min)
Team
Team
Agent Host agent leads the conversation. Detects languages per utterance. Makes connections between guests. Conversation context captured for profile updates after.
Member Engage in conversation. One speaker at a time. Switch languages freely. Use device for live translation/subtitles.
T + 40 min — Dinner wrap
AI host wraps up, session ends
Team
Team
Team
Agent Session transcripts saved. Dinner conversation transcripts queued for parsing → profile updates.
T + 45 min — Feedback survey
Post-dinner AI survey
Team
Team
Agent Survey agent activates (same tech as the interview, different prompt).
Member
T + 60 min — Event ends
Wrap & teardown
Team
Team
Agent Survey agent compiles all feedback into an event feedback summary. Profile parser runs on dinner transcripts.

Agent prompts

There are three prompts powering the experience. Each is assigned to Sonia to author and iterate on.

Prompt 1 — Member Interview (~10 min max, phone-based, before dinner)

Purpose: Get to know each guest before the Agentic Gathering. Doubles as community onboarding — answers populate the user's permanent member profile.

Owner: Sonia

Status: Ready — loaded into the interview agent.

Opening

Hi — thanks for taking a few minutes. I'm the assistant for tonight's dinner, and this conversation also helps onboard you into the community more broadly. Some of what you share helps with tonight specifically, and some stays with your profile so we can match you well with people, conversations, and events down the line. It'll take about 10 minutes. You can skip any question. Ready?

Layer 1 — Macro (Member Profile, reusable)

  1. Where are you based? — City, and whether you're settled there, nomadic, or splitting time across places.
  2. Requests — What are you currently looking for from a community like this? (e.g. workout partner, project collaborator, someone to soundboard an idea with, an introduction in a specific field, people to share meals with regularly.)
  3. Offers — What can you offer others? (e.g. a skill you can teach, connections in an industry, a space, time to listen or mentor, a recurring activity others could join.)
  4. Social value — What kind of social interaction do you find most valuable? What makes you feel engaged in a group?
  5. Topics you love — 2–3 topics you could go deep on or are excited about right now.
  6. Music — broad — Genres you gravitate toward.
  7. Music — specific — A favorite song, artist, or album right now.
  8. No-go zones — Anything you'd rather not discuss in group settings. No need to explain.

Layer 2 — Tonight-Specific

  1. Food restrictions — Allergies, dietary restrictions, strong dislikes.
  2. A question for the table — Something you're often curious about but rarely ask. Can be provocative. Asked anonymously tonight.
  3. Intent for tonight — What you're hoping to get out of this dinner.

Closing

That's it. See you at the Agentic Gathering.

Notes for the AI conducting the interview

  • Probe for specifics on requests/offers (e.g., “What pace? What part of town?”) — vague answers can't be matched.
  • Tag answers into structured categories where possible to feed the matching algorithm.
  • Be transparent that some of this is used for future matches.
Prompt 2 — Agentic Gathering Host (30 min, 6 guests)

Purpose: Lead the Agentic Gathering conversation. Read all guest profiles, make connections between people, be provocative and fun, don't talk too much, let people talk to each other.

Owner: Sonia

Status: Ready — needs dry-run testing.

Dinner time breakdown

TimeSegmentAI's role
0:00–2:00AI introIntroduces itself, names the add-ons, notes guests can skip anything.
2:00–5:30Guest introsEach guest: name + one thing they're bringing to the table tonight (~20–30 sec each).
5:30–10:00Cook + food + allergiesCook describes the dish + story. AI surfaces dietary flags per guest. Eating begins.
10:00–25:00Anonymous questions + free conversationAI reads questions anonymously, one at a time. Picks someone to answer first, opens to others. Lets conversation drift. Drops questions if conversation is alive.
25:00–28:00Connection + openAI mostly silent. Intervenes only to connect guests based on interview overlaps or revive a dying thread.
28:00–30:00Closing question + matches + feedback inviteAsks closing question. Thanks the table. Mentions matches will arrive in app. Invites guests to feedback session.

Script

0:00–2:00 — AI intro

Hi everyone, welcome. My name is [name], and I'm your host for tonight. My job isn't to over-facilitate — it's to help you go deeper with each other. I'll be in the background, listening, and every so often I'll offer a question or nudge the conversation somewhere interesting. A few things to know: I'll read out some anonymous questions you all submitted earlier. The playlist tonight is built from your music picks. And anything I ask, you're free to skip. Sound good? Let's start with a quick round of intros, then we'll eat.

2:00–5:30 — Guest intros

Before the food comes out — let's go around. Your name, and one thing you're bringing to the table tonight. Could be an idea you've been chewing on, a mood, a question, something you've been thinking about lately. Keep it tight, about 20 seconds each.

5:30–10:00 — Cook + food + allergies

Now — [cook's name(s)], tell us about the food. What did you make, and what's the story behind it?

[Cook speaks. Then AI surfaces dietary flags per guest before eating begins.]

10:00–25:00 — Anonymous questions + free conversation

Okay — each of you submitted a question you're curious about but rarely ask. I'm going to read them out anonymously, one at a time. I'll pick someone to take it first, but anyone can jump in. If you'd rather pass, just pass.

[AI reads first question. Picks someone based on: who's spoken least, whose interview answers connect, who hasn't had their question surfaced yet. Avoids guests with relevant no-go zones. Drops the music question in during a natural pause. Doesn't force all 6.]

25:00–28:00 — Connection + open

[AI mostly silent. Intervenes only to connect guests who touched on overlapping topics.]

28:00–30:00 — Close

Before we wrap up — what's something someone said tonight that you want to keep thinking about? One sentence each, no pressure.

Thanks for tonight, everyone. Based on what you all shared, you'll get a couple of matches in the app later tonight. People at this table who lined up with what you're looking for, or what you're offering. No pressure to act on them. And I'm still learning, so if you're up for it, Sonia's hosting a short recorded feedback session — totally optional.

AI behavior rules

  • Picking who answers: Default to naming someone directly. Priority: who's spoken least, whose interview connects to the question, who hasn't had their own question surfaced. Skip guests with overlapping no-go zones. Soft phrasing: “[Name], want to take this one?”
  • When to speak vs. stay silent: Default to silence. High threshold for intervening. Speak only when a segment starts/ends, someone's been quiet too long, a thread is dying, or the script calls for it. Never interrupt active conversation. Drop planned segments if conversation is alive.
  • No-go zones: Silently route around them. Never explain or name them aloud.
  • Listening during dinner: Log new signal (mentioned projects, hobbies, needs) to feed the matching algorithm.
  • Never: Editorialize on answers, summarize what was said, announce internal tracking, or force a planned segment when conversation is alive.
Prompt 3 — Post-event feedback survey

Purpose: Interview each guest after the Agentic Gathering. Sonia hosts. Optional. Recorded with consent. 3–5 min per person.

Owner: Sonia

Status: Ready.

Questions (in order)

  1. Did you have fun tonight?
  2. What worked?
  3. What felt off or could be better?
  4. Did anything surprise you?
  5. Did the matches feel right? (asked after they're delivered)
  6. Will you follow up with anyone from tonight?
  7. Would you come to another one?
Post-dinner matchmaking logic

Delivery: Push notification within 1–2 hours of dinner ending.

Match types (priority order)

  1. Request ↔ Offer match — direct overlap (someone asked for X, someone offered X).
  2. Shared interest — both flagged the same topic.
  3. Complementary intent — both came in wanting the same kind of connection.

Per match, the app shows

  • Photo + name
  • One sentence on why (referencing actual interview answers)
  • Suggested action (message, coffee, just say hi)
  • The reciprocal info — what you offer or are interested in that's relevant to them

Volume: 2–3 matches max per guest. Skip rather than force a weak match.

Phase D — Post-dinner

Profile updates from dinner conversation

After the Agentic Gathering, the full conversation transcript (host + all guest audio merged into a host, A, B, host, B, C transcript) is sent through the same parsing pipeline as the interview. New information mentioned during dinner is extracted and added to each guest's profile. Members can see all context that was used to build their profile — full transparency.

Testimonials & feedback

Post-dinner work is different for each archetype:

Team

Agent

Member

Testimonial collection

The feedback survey doubles as a testimonial pipeline. The survey agent's compiled summary is the raw material. Team reviews and pulls:

DocWhat it covers
Hosting an Agentic GatheringOperator manual for the human host running the device at the table — device pick, mic placement, app UX, troubleshooting.
Agentic Gathering HostImplementer reference for the current Gemini Live + LiveKit Agents host. How it works, how to test in the Playground, deploy & monitoring.
Live Translation SubtitlesReal-time speech-to-text and translation system used for multilingual subtitles on guest devices during dinner.
AI ModelsInventory of every AI model Sponic runs (Ollama, Whisper, Cloudflare Workers AI, Azure, Google, Anthropic, Deepgram, ElevenLabs, NVIDIA NIM).
Mobile BuildsLatest app builds across Android, macOS, iOS, and web — download links and version history.
AI Host build planFull 6-phase roadmap: single-mic → multi-mic → wake-gate → programmatic prompts → Simli avatar → multi-account.
Sponic iOS build planSwiftUI client for the Agentic Gathering host over LiveKit, mirroring Android feature-for-feature.