exec-ops-env / openenv.yaml
saaheerpurav's picture
first commit
1bb18a0
Raw
History Blame Contribute Delete
3.38 kB
name: ai-executive-ops-manager
version: "1.0.0"
description: >
Simulates a startup CEO's (Alex Rivera, NovaTech AI) operational day.
An AI agent must manage a realistic executive inbox under time pressure,
making triage decisions - reply, schedule, delegate, or ignore - to
handle conflicting priorities, cascading crises, and stakeholder demands.
Tests prioritization, urgency sensitivity, and decision quality.
entrypoint: inference.py
tasks:
- id: easy
name: Monday Morning Catchup
difficulty: easy
description: >
A manageable Monday inbox with 4 emails and clear priority separation.
Optimal play: approve production deploy (P5), handle enterprise contract (P4),
delegate onboarding (P3), ignore snack order (P1).
max_steps: 6
expected_score_range: [0.7, 1.0]
- id: medium
name: Investor Demo Day Prep
difficulty: medium
description: >
Series B investor demo day. 7 emails with genuine scheduling conflicts -
CTO architecture crisis competes with investor prep. Requires smart
delegation and scheduling around locked calendar slots.
max_steps: 8
expected_score_range: [0.45, 0.85]
- id: hard
name: Series B Crisis Day
difficulty: hard
description: >
Cascading crises: investor pulling term sheet, production DB corruption,
key engineer resignation, TechCrunch security breach article going live,
GDPR deadline, and more. 10 emails, 4 P5 items, 8 steps - ruthless
triage required. Perfect score is intentionally unachievable.
max_steps: 8
expected_score_range: [0.2, 0.75]
observation_space:
description: >
Dict with time (str), step (int), max_steps (int), steps_remaining (int),
inbox (list of Email objects), calendar (list of CalendarEvent objects),
goals (list of all Goal objects with completion status),
pending_goals (list of incomplete Goal objects), total_reward (float), done (bool).
email_fields:
- id: str
- sender: str
- subject: str
- body: str
- priority: "int 1-5 (1=minimal, 5=critical)"
- urgency: "float 0.0-1.0, decays by 0.15 each step"
- handled: bool
- action_taken: "str or null"
action_space:
type: discrete
actions:
- reply: "Personally respond to the email (best for high-priority items requiring CEO attention)"
- schedule: "Book a meeting slot for this item (good for relationship/strategic items)"
- delegate: "Assign to a team member (good for operational P1-P3 items)"
- ignore: "Skip this item (only appropriate for P1 items when critical items are unhandled)"
format: '{"type": "reply|schedule|delegate|ignore", "email_id": "<id>"}'
reward:
type: continuous
per_step: true
range: "[-0.3, 0.5] per step"
components:
- goal_completion: "+0.30 * (goal_priority / 5) for completing a goal"
- urgency_bonus: "+0.15 * urgency if urgency > 0.5 and not ignoring"
- efficiency: "+0.05 for delegating P1-P2 items"
- action_match: "+0.10 if action matches goal's required action"
- ignore_penalty: "-0.25 for ignoring P4-P5 emails"
- waste_penalty: "-0.10 for replying to P1 when P4+ unhandled"
- step_cost: "-0.02 per step (encourages efficiency)"
grader:
type: deterministic
score_range: [0.0, 1.0]
formula: "0.60 * goal_completion_rate + 0.25 * email_handling_rate + 0.15 * efficiency_bonus"