Interactive Demo

Watch a political agent think.

A US Senator evaluates a bill in real time — retrieving context, resolving conflicting pressures, and producing a structured decision with downstream event emissions. This is one turn of a 535-agent simulation.

Demo mode: agent reasoning is pre-computed so it runs without live infrastructure. The logic, JSON schemas, and output format are production-identical.

Political agent — US Senate, Florida

Republican Senate · FL schema v1.0 · completeness 0.91

Sen. James R. Holden (R-FL)

agent_id: senate_fl_1_holden · bioguide_id: H001234 · last_updated: 2026-04-01

Economic ideology

0.78

Party loyalty

0.84

Reelection priority

0.72

Donor responsiveness

0.68

Media sensitivity

0.55

Risk tolerance

0.40

llm_persona_prompt — injected as Bedrock system prompt on every agent turn A conservative Florida senator with strong party loyalty, substantial healthcare and insurance industry donor exposure, and moderate reelection vulnerability. Responds predictably to leadership signals on most votes. Will consider cross-party movement when district healthcare access salience is high and primary risk is low. Rarely a first-mover in coalition formation. Avoids positions that directly conflict with his top donor bloc.

Select a bill scenario

tick 4

Active simulation event

Agent turn protocol

Retrieve → Evaluate → Decide → Persist

Bedrock Haiku — context package assembly (1,800 token budget)

Implementation reference

In production each agent turn executes in under 4 seconds. At FLOOR_VOTE_CALLED scale, up to 535 agent turns run simultaneously via AWS Step Functions parallel Map state — each retrieving its own context, evaluating against its own calibrated persona, and emitting typed events that propagate causally through the simulation until a terminal institutional state is reached.

← Main site Platform architecture Request investor package