civitas: Instructions file for Codex

AGENTS.md

Project

Civitas — Autonomous Food Procurement through Agent Parliament and Agent Jury

Civitas is a multi-agent autonomous procurement system based on Problem Statement 5 of the AgentXcelerate challenge.

The official problem asks the system to determine what food to procure, how much to procure, from which suppliers, and how inventory should be distributed across locations while considering demand, inventory, supplier lead times, pricing, perishability, and location requirements.

Civitas goes beyond this baseline by introducing:

1. Agent Parliament — autonomous agents with conflicting objectives negotiate a procurement plan.
2. Agent Jury — evaluates whether the reasoning and evidence behind the agreement deserve to be trusted.
3. Adaptive Replanning — a weak Jury result sends the system back into investigation and negotiation instead of allowing immediate execution.

---

1. Core Product Principle

Don't just make agents agree. Make sure their agreement deserves to be trusted.

The system must distinguish:

Agent Agreement
        ≠
Independent Evidence
        ≠
Correct Decision

A proposal supported by six agents can still be based on one underlying source or one propagated assumption.

Civitas therefore treats evidence lineage as a first-class object.

---

2. Architecture

The primary control loop is:

USER GOAL
    ↓
PLANNER
    ↓
PARLIAMENT
    ↓
INVESTIGATION THROUGH MCP
    ↓
COMPETING PROPOSALS
    ↓
NEGOTIATION
    ↓
CANDIDATE PLAN
    ↓
JURY
    ↓
┌───────────────┐
│               │
▼               ▼
APPROVE      INVESTIGATE
│               │
▼               ▼
EXECUTE       REPLAN
                │
                └──────→ PARLIAMENT

The Jury is therefore part of the autonomous planning loop.

It is not a final dashboard or post-hoc evaluator.

The primary product interface is conversational:

OPERATOR → CODEX → CIVITAS MCP SERVER → CONTROL LOOP
                                      → OPERATIONAL MCP PROVIDERS