Crisora

landing page
multi-agent PR crisis matrix
agents
baseline agent solved the issue
crisis snerio flow
final results
architectural diagram

Inspiration

When a corporate crisis hits—whether it's a 100k-record data breach or a critical cloud outage—PR and Legal teams rarely see eye to eye. Legal prioritizes liability defense and minimal exposure, while PR demands transparency and customer trust. Standard single-agent LLM prompts try to compromise, yielding lukewarm responses that satisfy neither side.

We built Crisora for Track 3 (Agent Society) to solve this by deploying an asynchronous multi-agent boardroom negotiation system that turns domain conflict into an asset rather than a bottleneck.

What it does

Crisora simulates a multi-phase corporate boardroom crisis response:

Single-Agent Baseline: Runs a control prompt to establish baseline performance.
Parallel Adversarial Drafts: Simultaneously deploys a Radical Legal Council Agent (liability-focused) and an Empathic PR Brand Guardian Agent (trust-focused).
Arbitrated Consensus: A Boardroom Arbitrator Agent synthesizes both opposing perspectives, resolves policy contradictions, and outputs a final strategy along with a RESOLVED status and a quantitative utility score.

How we built it

LLM Engine: qwen-max via Alibaba Cloud DashScope for complex reasoning, role-playing, and strategy synthesis.
Serverless Backend: Hosted on Alibaba Cloud Function Compute (FC) to handle low-latency orchestration and parallel agent streams.
Frontend UI: Built with React + Vite featuring a glassmorphism dashboard that highlights live agent thought logs and performance deltas (+10% utility lift).

Agents

The Legal Agent drafts a strategy prioritizing compliance and regulatory obligations.
The PR Agent simultaneously drafts a strategy focused on tone and brand sentiment.
The Arbitrator Agent evaluates both streams, resolves conflicting directives, and synthesizes a unified crisis mitigation roadmap.

Challenges we ran into

Balancing opposing prompt constraints without causing model convergence was tough. We tuned system prompts carefully so Legal remained uncompromisingly defensive and PR remained deeply empathetic before handing both outputs to the Arbitrator.

Accomplishments that we're proud of

Achieving a 10% utility score increase over single-agent baselines and demonstrating that specialized agent societies produce far more actionable, legally sound, and empathetic crisis plans.