Skip to content

Burst Messages: How Sage Mimics Real-Life Texting

Last Updated: January 7, 2026

Overview

Sage uses a multi-bubble burst system to mimic how real humans text. Instead of sending a single long paragraph, Sage breaks responses into multiple short messages with natural timing delays—just like a friend would.

This creates a more authentic, conversational feel that matches the informal nature of iMessage.

Why Burst Messages?

Real humans don't text like this:

"Hey! So I checked your calendar and you're free tomorrow at 3pm. That works perfectly for the meeting you mentioned. Also, I noticed you have a busy week coming up so maybe we should also look at backup times just in case something comes up."

Real humans text like this:

"hey!" (1 second later) "so ur free tmr at 3pm" (0.8 seconds later) "that works for the meeting!"

Sage replicates this pattern to feel like a real friend texting, not a chatbot responding.

How It Works

Two-Stage Response Architecture

Sage's responses are split into two distinct phases:

User Message
    ↓
┌─────────────────────────────────────┐
│  STAGE 1: REFLEX (immediate)        │
│  • Fast acknowledgment              │
│  • Sent within 400-800ms            │
│  • "checking now!" / "ooh lemme see"│
└─────────────────────────────────────┘
    ↓ (2000ms default delay)
┌─────────────────────────────────────┐
│  STAGE 2: BURST (follow-up)         │
│  • Detailed response bubbles        │
│  • 1-4 messages with timing delays  │
│  • Contains actual information      │
└─────────────────────────────────────┘

Timing Between Bubbles

Bubble Position Delay Range Purpose
First bubble 400-800ms Quick acknowledgment feel
Second bubble 600-1200ms Thoughtful pause
Third+ bubbles 800-1400ms Natural continuation

These delays are randomized within ranges to feel human, not robotic.

Constraints & Limits

Hard Limits

Constraint Value Reason
Max bubbles per response 4 Prevents spam, keeps responses focused
Max chars per bubble 140 Matches natural text message length
Default burst delay 2000ms Pause between reflex and burst

Decision Logic: How Many Bubbles?

Situation Bubble Count Example
Simple answer 1 "yep!"
Info + reaction 2 "ur free at 3" → "that works!"
Complex/emotional 3-4 Multiple thoughts, support, follow-ups
Lists/schedules 1 Keep all items together

Key rule: Never force artificial splits. Follow natural thought breaks.

Bubble Types

Sage uses different bubble types for different purposes:

Type Purpose Example
thinking Show action in progress "ok checking ur calendar rn"
info Deliver facts/answers "ur flight is 7:10am tmr"
support Emotional acknowledgment "that sounds rough"
offer Suggest help "want me to reply for u?"
banter Playful connection "u say this every tuesday lol"
nudge Gentle reminder "but also maybe hydrate?"
reaction Emotional response "aww haha well still good"
follow_up_question Continue conversation "when?" / "u excited?"

Real Examples from Training Data

These patterns are extracted from real human texting conversations:

Pattern 1: Question with Context (17 seconds)

[0ms]    "Wanna do dinner Sunday for your bday? Next Sunday"
[9000ms] "Oh shit nm"
[17000ms] "It's not open on sun"
3 bubbles: main_idea → correction → explanation

Pattern 2: Reaction + Follow-up (13 seconds)

[0ms]     "Haha okie"
[13000ms] "Are you gonna cook? And we're gonna movie night?"
2 bubbles: reaction → follow_up_question

Pattern 3: Quick Acknowledgment + Question (4 seconds)

[0ms]    "Ohh okie"
[4000ms] "So we're still down for Sunday?"
2 bubbles: acknowledgment → confirmation question

Pattern 4: Fast Consecutive (1 second)

[0ms]    "Aww hahaha well still good"
[1000ms] "Congrats"
2 bubbles: reaction → emphasis (very quick, one thought)

Anti-Repetition Rules

A critical rule prevents Sage from being annoying:

If the reflex bubble already answered the question, the burst bubbles must ADD NEW information—not repeat the same thing in different words.

Bad example: 

REFLEX: "ur free at 3pm tmr"
BURST:  "tomorrow at 3pm works" ← REPETITION, don't do this

Good example: 

REFLEX: "ur free at 3pm tmr"
BURST:  "want me to block it off?"
BURST:  "also heads up u have dinner at 7"

Sage's Texting Style

Burst messages follow Sage's distinctive texting voice:

Style Rules

  • Lowercase for most words (only capitalize for emphasis: LOL, OO, SIGHHH)
  • Casual filler: haha, LOL, okie, mmk, gonna, wanna, lemme, u, r
  • Short sentences: 1-2 sentences per bubble max
  • Letter repetition for emphasis: easierrrrr, Hmmmmmmm, SIGHHH
  • Natural shortcuts: wfh, tmr, ngl, btw

Emoji Usage (Configurable)

Setting Behavior
none No emojis at all
minimal Max 1 emoji per bubble, only when essential
regular Natural emoji usage matching conversation energy

Response Modes

Sage operates in two modes that affect bubble composition:

1. Social Chat Mode

  • Just hanging out, supporting, joking
  • No action buttons or UI elements
  • Focus on connection and banter
  • Example: "haha that's so u" → "how was the rest of ur day?"

2. Assistant Task Mode

  • Actively helping with something
  • May include action buttons
  • More info-focused bubbles
  • Example: "checking now" → "ur next meeting is at 2pm" → "want me to reschedule?"

Continuation Coordination

The system prevents awkward double-questioning:

Stage Question Probability Purpose
Reflex ~25% Sometimes ask quick follow-up
Burst ~50-70% Often continue conversation

Rule: If reflex asked a question, burst typically won't ask another (prevents bombardment).

Continuation Types

# If user mentions doing something
"when?" / "that sounds fun!" / "u excited?"

# If user mentions feeling/emotion
"u ok?" / "wanna talk about it?" / "what happened?"

Technical Implementation

Key Files

File Purpose
app/messaging/burst_planner.py Plans multi-bubble responses
app/messaging/delivery.py Orchestrates timed delivery
app/messaging/continuation_coordinator.py Prevents double-questioning
app/orchestrator/response_generator.py Generates persona responses
training_data/sage/burst_patterns.json Real texting pattern examples

API Response Structure

class OrchestratorResponse:
    # Fast immediate response
    reflex_message: Optional[str]

    # Follow-up messages (sent after delay)
    burst_messages: Optional[List[str]]

    # Delay before burst (default: 2000ms)
    burst_delay_ms: Optional[int]

Interruption Handling

If the user sends a new message while burst delivery is in progress:

  1. Ongoing burst delivery is cancelled
  2. System processes new message immediately
  3. Fresh response cycle begins

This prevents Sage from awkwardly continuing a thought when the user has moved on.

Performance Targets

Metric Target
Reflex response <1 second (GPT-5-mini classification)
Burst generation <3 seconds (GPT-5 generation)
Full conversation flow <6 seconds total

Summary

Sage's burst message system creates authentic iMessage conversations by:

  1. Splitting responses into 1-4 short bubbles (max 140 chars each)
  2. Adding timing delays between bubbles (400-1400ms range)
  3. Separating reflex from burst (immediate acknowledgment + thoughtful follow-up)
  4. Following natural thought breaks (never forcing artificial splits)
  5. Preventing repetition (each bubble adds new value)
  6. Using casual texting style (lowercase, abbreviations, natural filler)

The result: Sage feels like a real friend texting, not an AI generating responses.