Cicero

Inspiration

Local campaigns face the same challenges as national campaigns but without the staff, budget, or infrastructure.

A congressional campaign can afford pollsters, field directors, communications teams, and analysts who spend their days tracking what voters care about and planning outreach accordingly. A local race often runs on a handful of volunteers and limited resources. Important community issues can emerge and disappear before a campaign has time to react.

We asked a simple question:

What if every local campaign had access to an AI field team?

Rather than building another chatbot that waits for instructions, we wanted to build a system that proactively identifies opportunities for community engagement and recommends what a campaign should do next.

The result is Cicero: a multi-agent civic intelligence platform that helps campaigns discover local issues, engage communities, and prioritize outreach efforts.

What It Does

Cicero transforms local news, turnout data, and weather forecasts into a ranked slate of community outreach opportunities.

A user provides only:

A district or region
A planning horizon

From there, Cicero works autonomously.

The system:

Scans local news and community signals.
Identifies emerging issues affecting different communities.
Generates outreach events tied to those issues.
Pulls local weather forecasts for each proposed event.
Uses turnout data to identify target voter groups.
Prioritizes opportunities based on impact and feasibility.

The result is not another analytics dashboard.

It is a set of recommended actions:

"Here are the five most valuable events your campaign should run next."

Each recommendation includes a venue suggestion, weather-aware event format, target audience, talking points, rationale, and ready-to-use outreach content.

How We Built It

Cicero is powered by a team of cooperating AI agents, each responsible for a specific part of the planning process.

Issue Scout

The Issue Scout analyzes local information sources and identifies the most important issues affecting different parts of a district.

It produces structured issue reports containing:

Issue title
Geographic area
Supporting sources
Summary
Salience score

Event Architects

For every issue identified, Cicero launches an Event Architect.

These agents run independently and in parallel.

Each Event Architect:

Retrieves local weather forecasts using Jua
Queries precinct-level turnout information
Identifies relevant voter segments
Chooses an appropriate venue and event format
Generates issue-specific talking points
Creates draft outreach materials

This allows the system to evaluate multiple opportunities simultaneously.

Slate Strategist

The Slate Strategist receives all event recommendations and determines which opportunities deserve attention first.

Events are ranked using factors such as:

Issue importance
Community relevance
Turnout opportunity
Operational feasibility

The output becomes a prioritized action plan rather than a collection of disconnected insights.

Orchestrator

An orchestration layer coordinates the entire workflow:

Issue Scout → Event Architects → Slate Strategist

It manages parallel execution, structured outputs, caching, and fault tolerance while assembling the final ranked slate.

Challenges We Faced

Generating Actions Instead of Summaries

Many AI systems are good at summarizing information.

Generating useful next steps is far more difficult.

Our biggest challenge was transforming local issues into concrete, location-specific recommendations that campaign organizers could immediately act upon.

Multi-Agent Coordination

Because multiple agents collaborate to produce a final recommendation, maintaining consistent communication between agents was critical.

We addressed this by defining shared schemas and structured outputs that every agent follows.

Reliability During Live Execution

Hackathon demos can fail because a single API call fails.

To make the system resilient, each Event Architect operates independently. If one weather request or data query fails, the rest of the slate can still be generated successfully.

Maintaining Realism

We wanted recommendations to be grounded in real communities rather than hypothetical scenarios.

Balancing real-world data, practical event planning, and hackathon time constraints required careful scope management and strong architectural boundaries.

What We Learned

Building Cicero reinforced an important lesson:

The most valuable AI systems don't just answer questions—they surface opportunities and recommend actions.

We also learned that:

Structured outputs dramatically improve agent reliability.
Specialized agents are easier to reason about than monolithic systems.
Parallel agent architectures allow teams to build faster during hackathons.
Combining multiple real-world signals produces significantly better recommendations than relying on a language model alone.

Most importantly, we learned that autonomy is not measured by how much text an AI can generate. It is measured by how much useful work it can perform before requiring human intervention.