Prism Swarm 🔮

What is Prism?

Prism is an autonomous multi-agent system that thinks like a team.

Give it a question. It assembles the right agents. They debate. They disagree. They reach consensus. They deliver one answer.

---

The Challenge

SQL generation isn't a prompt engineering problem.
It's a reasoning problem.

Real-world queries demand:

• Understanding business context, not just table names
• Navigating complex schemas with 100+ tables
• Reasoning through temporal logic, aggregations, and edge cases
• Making decisions when the path forward isn't clear

How Prism Works

Multi-Agent Architecture

Specialized agents with distinct expertise:

Autonomous Consensus

When agents disagree, they negotiate.
When the path is uncertain, they explore.
When mistakes happen, they learn.

The system decides. Not the engineer.

One Question, One Answer

The agents reach consensus on a single solution—or they iterate until they do.

---

Why It Matters

Most "agentic" systems are still just LLMs with function calls.

Prism is different:

• Agents have goals, not just instructions
• Conflicts are resolved through reasoning, not heuristics
• The system adapts, without retraining

---

Results

We submitted Prism to SPIDER 2.0 (SNOW Track):

• Claude Sonnet 4.5 as the reasoning substrate
• Zero manual intervention – agents decide, humans observe
• Complete decision traces – every negotiation, every conflict, every resolution

Every query answered through agent consensus, not engineering hacks.

Score

Evaluated against the Spider 2.0 benchmark: https://spider2-sql.github.io/

<img width="789" height="742" alt="Screenshot 2026-01-05 at 2 26 24 PM" src="https://github.com/user-attachments/assets/546dacc8-28fe-43ed-9d2e-b00121b2a6dc" />

---

What's Next

Prism treats query generation as a multi-player game where agents cooperate toward a shared objective, not a state machine where transitions are hardcoded.

Coming soon.

The code stays proprietary. The ideas don't.

---

Technical Details

Model

• Claude Sonnet 4.5 (Anthropic)
• Temperature: 0.0 (deterministic)
• Context: 200K tokens

---

The Evolution of Agentic Systems

Most research optimizes individual agent capability.

We optimized multi-agent coordination.

It's not about building a better player. It's about building a better game—where cooperation yields better outcomes than competition, and consensus emerges through incentive alignment, not voting.

---

Team

Anshul Chauhan · anshul1.chauhan@paytm.com
Soham Acharya · soham.acharya@paytm.com

Paytm

---

Paper coming soon.

prism

Agent Swarm for Automated SQL Generation