I built a 12-state accountability layer for AI agents

Описание и применение

Создана система COGEXT, отслеживающая обещания ИИ‑агентов через 12 состояний, обеспечивающая атомарные переходы и дедупликацию.

Задача

AI agents generate commitments but fail to track or fulfill them, leading to forgotten promises and unreliable behavior.

Результат

COGEXT provides a robust, deterministic accountability layer that ensures commitments are tracked, deduplicated, and verified, reducing missed deadlines and improving agent reliability.

Рабочий процесс

• 1. Receive agent output
• 2. Classify text for genuine commitments (confidence ≥0.5)
• 3. Extract commitment fields with Groq llama-3.3-70b in JSON mode
• 4. Validate extraction against Pydantic schema
• 5. Normalize deadlines deterministically
• 6. Compute idempotency key and insert commitment via PostgreSQL RPC
• 7. Track state transitions through 12 states using a single RPC function
• 8. Aggregate evidence to determine fulfillment
• 9. Emit immutable events and audit logs

Доказательства

All state mutations go through a single PostgreSQL RPC function. Not direct UPDATEs from application code

Источник

We validate everything against Pydantic schemas before touching the database. No exceptions.

Источник

The hour window is deliberate: it deduplicates within a session while still allowing the same promise to be legitimately re-made days later.

Источник