An AI assistant that searches apartments, books flights, manages email, watches prices, and remembers everything — all with trust-gated safety. Send a voice note on Telegram. Get results before your coffee is ready.
28 compound skills turn multi-step workflows into single requests. Just ask.
You spend hours each week jumping between Zillow, Google Flights, Gmail, and Google Calendar to handle tasks that should take seconds. DAT Agent collapses those workflows into a single conversational request — delivered through Telegram, Slack, or Teams.
You (Telegram voice note):
"Find me a 2-bedroom apartment
in Dublin, Ohio under $1,800
and email the top 3 to my wife"
DAT Agent:
1. search_apartments
-> Zillow API (structured data)
-> 7 listings found
-> Filtered to Dublin, OH only
2. send_formatted_email
-> HITL approval requested
-> [Approve] [Deny]
-> You tap Approve
3. Result delivered to Telegram:
"Sent 3 apartments to Sarah.
Best deal: $1,450/mo, 2bd/1ba
at 5820 Blazer Pkwy, Dublin"
Total time: 14 seconds
Total iterations: 3
LLM calls: 2 (Haiku cached)Your agent starts cautious and earns its way to full autonomy. You stay in control of the big decisions.
Most AI assistants are all-or-nothing: either they can do everything or nothing. DAT introduces graduated trust. Your agent starts in a read-only sandbox and earns capabilities through demonstrated reliability. High-risk actions like sending emails or filling forms always require your explicit approval — delivered as a tap-to-approve button right in your chat.
Trust-Gated Capability Ladder
==============================
STRICT Sandbox (trust 0-30)
read_file list_directory
search_files
-> 3 tools, 3 max iterations
ADAPTIVE Sandbox (trust 30-70)
+ web_search + fetch_url
+ execute_command (read-only)
+ browse_url + extract_data
+ recall + remember
+ All 28 compound skills
-> 34 tools, 20 max iterations
OPEN Sandbox (trust 70-100)
+ send_email [approval required]
+ click_element [approval required]
+ fill_form [approval required]
+ delegate_task [approval required]
+ build_skill [approval required]
-> 45 tools, 25 max iterations
Every high-risk action shows:
Tool: send_formatted_email
To: [email protected]
Subject: Top 3 Dublin Apartments
[Approve] [Deny]Your agent remembers your preferences, past searches, and context across every conversation.
Generic AI assistants forget everything between sessions. DAT Agent maintains two layers of memory: short-term conversation context (your last 5 tasks) and long-term semantic memory powered by pgvector. When you say "send that to my wife," the agent knows exactly what "that" refers to and who your wife is — because it remembers.
The result is an agent that gets better the more you use it. It learns your preferences, remembers your contacts, and anticipates your needs — all within a trust-gated, privacy-preserving framework.
Memory Architecture
==============================
Short-Term (Redis, 24h TTL):
dat:agent:memory:telegram:874512
-> Last 5 task summaries
-> "Found 7 apartments in Dublin"
-> "Emailed Sarah the top 3"
-> "Booked flight to Cancun"
Long-Term (pgvector, persistent):
agent_memories table
-> 1536-dim embeddings
-> Hybrid search score:
0.7 * cosine_similarity
+ 0.3 * recency_decay(14d)
+ importance_boost
Auto-Indexed Categories:
email, web, browser, file,
delegation, general
Follow-Up Example:
You: "Send that to my wife"
Agent recalls:
Memory #1: "Sarah = wife"
Memory #2: "Last task: apartments"
-> Sends apartment list to Sarah
-> No clarification neededStart with a free account. Send your first voice note on Telegram. Watch your agent earn its capabilities in real time.