Skip to content

OzWizards

DriftPA trains LLM agents to be reliable personal executive assistants in a world that breaks without warning.

Code Demo Video Cerebral Valley

Core trick

GRPO trains a personal-assistant agent under injected schema drift, expiring tasks, and irreversible booking actions in a simulator.

Tags

agentsproductivity

Description

DriftPA trains LLM agents to be reliable personal executive assistants in a world that breaks without warning. Real AI assistants fail in production because APIs change schema mid-session, tasks expire while the agent deliberates, and wrong bookings or emails can't be undone. DriftPA is an RL environment that simulates exactly these four failure modes simultaneously: schema drift (field names change mid-episode), time pressure (urgent tasks expire), irreversible actions (book/reply/cancel can't be undone), and policy drift (cancellation windows tighten). An untrained agent scores -9.55 mean reward — it uses stale API fields, misses the boss's urgent email, double-books dinner, and triggers cascade failures. A GRPO-trained agent learns to call list_tools() after drift, prioritize by urgency, and commit to irreversible actions only when safe. Optimal score: +22. Untrained mean: -9.55. That 31-point gap is the problem we solve.
rajashekarcs2023/openenv-RL★ 00 forksPythonCreated Mar 8, 2026 · last push Mar 10, 2026

README excerpt

# DriftPA — Personal Assistant RL Environment

**Trains LLM agents to handle real-world chaos: schema drift, policy changes, time pressure, and irreversible decisions.**

OpenEnv Hackathon | Track 3.2 (Personalized Tasks) + Patronus AI (Consumer Workflows)

---

## The Problem

Real AI assistants break in production because:
1. **APIs change schema mid-session** — field names rename without warning
2. **Tasks expire** — the boss's urgent email can't wait forever
3. **Wrong actions can't be undone** — a bad booking is a bad booking
4. **Policies tighten** — cancellation windows shrink after drift

No existing RL environment trains agents against all four simultaneously.

## DriftPA

A personal executive assistant environment where the agent must resolve cascading real-life conflicts — calendar clashes, urgent emails, dinner bookings, ride scheduling — while four failure modes fire without warning.

| Agent | Mean Episode Reward |
|---|---|
| Untrained baseline | −9.55 (min −22) |
| Optimal sequence | +22.0 |

## Four Novel Mechanics

1. **Schema Drift** — `party_size → guests`, `pickup_time → eta_minutes`. Agent must call `list_tools()` to discover new schema or get penalised for stale fields.
2. **Time Pressure** — Boss email expires at step 4. Missing it triggers a cascade.
3. **Irreversible Actions** — `reply_message`, `book_restaurant`, `book_ride`, `cancel_booking` cannot be undone.
4. **Policy Drift** — Cancellation window tightens from 2hr → 4hr post-drift.

## Figure:

Team