tl  tr
  Home | Tutorials | Articles | Videos | Products | Tools | Search
Interviews | Open Source | Tag Cloud | Follow Us | Bookmark | Contact   
 Generative AI > Google Gemini API > ADK Multi-Agent Debate Pattern

ADK Multi-Agent Debate Pattern

Author: Venkata Sudhakar

The debate pattern uses multiple agents to improve decision quality. One agent argues for a position, another argues against it, and a judge agent evaluates both arguments and delivers a verdict. This is particularly useful for high-stakes decisions where a single agent might miss important counterarguments. ShopMax India uses it for supplier selection decisions involving large procurement contracts.

The three roles are: advocate (argues in favour), critic (argues against), and judge (evaluates both and decides). Each agent sees the same input but has a different instruction that biases it toward its role. The judge agent receives both outputs before deciding.

The below example shows how ShopMax India uses the debate pattern to evaluate whether to approve a new supplier contract.


Running the three-agent debate and collecting the verdict,


It gives the following output,

=== ADVOCATE ===
- Price is 11% below market rate, saving Rs 40,000 per month
- High volume commitment secures dedicated production slots
- Shenzhen supplier has direct factory access

=== CRITIC ===
- 100% advance with no quality guarantee is unacceptable risk
- No recourse for defective units once payment made
- 45-day lead time creates inventory gaps

=== JUDGE ===
DECISION: REJECT (Confidence: 82%)
The price advantage does not justify 100% advance payment with no
quality guarantee. Counter-propose: 30% advance, 70% on delivery,
with quality inspection clause.

The debate pattern scales well - add domain-specific agents like a legal reviewer or a finance analyst as additional participants. The judge agent receives all perspectives before deciding. For ShopMax India, this pattern reduced procurement errors by 40% compared to single-agent decisions on contracts above Rs 10 lakhs.


 
  


  
bl  br