← All Cookbooks
AutoGenAdvanced30 min
AutoGen + HatiData: State Branch & Merge
Use HatiData branching to let AutoGen agent groups explore multiple solution paths simultaneously, then merge the best outcome.
What You'll Build
An AutoGen multi-agent conversation that uses branching to explore multiple solution paths simultaneously.
Prerequisites
$pip install autogen-agentchat hatidata-agent
$hati init
$OpenAI API key
Architecture
┌──────────────┐ ┌──────────────────┐
│ AutoGen │───▶│ HatiData │
│ GroupChat │ │ Branching │
└──────────────┘ └────────┬─────────┘
┌───────┼───────┐
▼ ▼ ▼
Branch A Main Branch B
(plan 1) (plan 2)
└───────┬───────┘
Merge bestKey Concepts
- ●Parallel exploration: create multiple branches to test competing strategies without risk to the main state
- ●Schema isolation: each branch is a separate isolated schema, so writes in one branch never leak to another
- ●Merge strategies: branch_wins overwrites main, main_wins keeps main, manual does row-level conflict resolution
- ●Automatic cleanup: discarded branches are garbage-collected, reclaiming disk space and schema entries
- ●Zero-copy reads: branches start as views over main state, only materializing tables on first write
Step-by-Step Implementation
1
Install Dependencies
Install AutoGen and the HatiData agent SDK.
Bash
pip install autogen-agentchat hatidata-agent
hati initExpected Output
HatiData initialized successfully.
Proxy running on localhost:5439Note: Branching creates isolated schemas. Each branch can have independent writes without affecting main state.
2
Create Baseline State
Set up the main data state that agents will branch from.
Python
from hatidata_agent import HatiDataAgent
hati = HatiDataAgent(host="localhost", port=5439, agent_id="autogen-planner")
# Create deployment configurations table
hati.execute("""
CREATE TABLE deploy_config (
service VARCHAR, replicas INTEGER,
cpu_limit VARCHAR, memory_limit VARCHAR,
region VARCHAR
)
""")
hati.execute("""
INSERT INTO deploy_config VALUES
('api-gateway', 3, '500m', '512Mi', 'us-east-1'),
('ml-inference', 2, '2000m', '4Gi', 'us-east-1'),
('data-pipeline', 1, '1000m', '2Gi', 'us-east-1')
""")
result = hati.query("SELECT service, replicas, region FROM deploy_config")
print("Baseline deployment config:")
for r in result:
print(f" {r['service']}: {r['replicas']} replicas in {r['region']}")Expected Output
Baseline deployment config:
api-gateway: 3 replicas in us-east-1
ml-inference: 2 replicas in us-east-1
data-pipeline: 1 replicas in us-east-13
Create Parallel Exploration Branches
Create two branches for different scaling strategies.
Python
# Branch A: horizontal scaling (more replicas, same resources)
hati.execute("SELECT branch_create('horizontal-scale')")
hati.execute("""
SELECT branch_query(
'horizontal-scale',
'UPDATE deploy_config SET replicas = replicas * 3'
)
""")
hati.execute("""
SELECT branch_query(
'horizontal-scale',
'INSERT INTO deploy_config VALUES (''cache-layer'', 6, ''250m'', ''256Mi'', ''us-east-1'')'
)
""")
# Branch B: vertical scaling (more resources, multi-region)
hati.execute("SELECT branch_create('vertical-scale')")
hati.execute("""
SELECT branch_query(
'vertical-scale',
'UPDATE deploy_config SET cpu_limit = ''4000m'', memory_limit = ''8Gi'' WHERE service = ''ml-inference'''
)
""")
hati.execute("""
SELECT branch_query(
'vertical-scale',
'INSERT INTO deploy_config VALUES (''ml-inference'', 2, ''4000m'', ''8Gi'', ''eu-west-1'')'
)
""")
print("Created branches: horizontal-scale, vertical-scale")Expected Output
Created branches: horizontal-scale, vertical-scaleNote: Each branch gets its own isolated schema. Writes in one branch never affect the other or main.
4
Compare Branch Outcomes
Query both branches to compare the two scaling strategies.
Python
for branch in ['horizontal-scale', 'vertical-scale']:
result = hati.query(f"""
SELECT branch_query(
'{branch}',
'SELECT service, replicas, cpu_limit, memory_limit, region FROM deploy_config ORDER BY service'
)
""")
total_replicas = sum(int(r['replicas']) for r in result)
print(f"\n=== {branch} ===")
for r in result:
print(f" {r['service']}: {r['replicas']}x ({r['cpu_limit']} CPU, {r['memory_limit']} RAM) in {r['region']}")
print(f" Total replicas: {total_replicas}")
# Main is unchanged
main = hati.query("SELECT SUM(replicas) AS total FROM deploy_config")
print(f"\nMain state total replicas: {main[0]['total']} (unchanged)")Expected Output
=== horizontal-scale ===
api-gateway: 9x (500m CPU, 512Mi RAM) in us-east-1
cache-layer: 6x (250m CPU, 256Mi RAM) in us-east-1
data-pipeline: 3x (1000m CPU, 2Gi RAM) in us-east-1
ml-inference: 6x (2000m CPU, 4Gi RAM) in us-east-1
Total replicas: 24
=== vertical-scale ===
api-gateway: 3x (500m CPU, 512Mi RAM) in us-east-1
data-pipeline: 1x (1000m CPU, 2Gi RAM) in us-east-1
ml-inference: 2x (4000m CPU, 8Gi RAM) in us-east-1
ml-inference: 2x (4000m CPU, 8Gi RAM) in eu-west-1
Total replicas: 8
Main state total replicas: 6 (unchanged)5
Merge Winner and Clean Up
Merge the chosen strategy into main state and discard the other branch.
Python
# Choose vertical scaling (fewer replicas, more efficient)
hati.execute("SELECT branch_merge('vertical-scale', 'branch_wins')")
hati.execute("SELECT branch_discard('horizontal-scale')")
print("Merged vertical-scale, discarded horizontal-scale")
# Verify final state
result = hati.query("SELECT service, replicas, cpu_limit, region FROM deploy_config ORDER BY service")
print("\nFinal deployment config:")
for r in result:
print(f" {r['service']}: {r['replicas']}x ({r['cpu_limit']} CPU) in {r['region']}")
# Verify no branches remain
branches = hati.query("SELECT branch_list()")
print(f"\nActive branches: {len(branches)}")Expected Output
Merged vertical-scale, discarded horizontal-scale
Final deployment config:
api-gateway: 3x (500m CPU) in us-east-1
data-pipeline: 1x (1000m CPU) in us-east-1
ml-inference: 2x (4000m CPU) in us-east-1
ml-inference: 2x (4000m CPU) in eu-west-1
Active branches: 0Note: After merging, the branch schema is cleaned up by the garbage collector. Active branch count should return to 0.