Back to Artifacts
Collaborative Era

The HAP Plan

The primary requirements artifact for designing collaborative human-agent systems.

The Approach

The HAP Plan (Human-Agent Pair Plan) is designed for systems where humans and AI agents work together as collaborative partners. Unlike User Stories or Agent Stories, the HAP Plan treats the pair itself as the unit of analysis—capturing not just what each party does, but how they work together, build trust, and evolve over time.

This artifact recognizes that in collaborative systems, capabilities are emergent: the pair can accomplish things neither the human nor the agent could do alone. It also acknowledges that the relationship changes—early interactions require more human oversight, while mature partnerships allow for greater agent autonomy.

The HAP Plan introduces concepts like confidence gates (milestones that unlock reduced oversight), guardrails (non-negotiable boundaries), and feedback loops (mechanisms for continuous improvement). It's the artifact of choice when designing AI assistants, copilots, collaborative tools, and any system where human judgment and agent capability must work in harmony.

Artifact Outline

1 Pair Goal & Setup

Shared Objective

The outcome the human-agent pair is trying to achieve together. Framed as a joint goal, not separate tasks.

Role Allocation

Initial division of responsibilities. What the human handles vs. what the agent handles at the start.

Shared KPIs

Metrics that measure pair success, not individual performance. Joint accountability indicators.

Tooling & Context

Systems, data sources, and interfaces the pair uses. The shared workspace for collaboration.

2 Gates & Guardrails

Confidence Gates

Measurable milestones that trigger reduced human oversight. Defined by accuracy, override frequency, or domain-specific metrics.

Delegation Triggers

Conditions under which the human "lets go" of specific tasks. The human's decision to trust, not just agent competence.

Hard Guardrails

Non-negotiable boundaries. Actions the agent must never take, regardless of confidence or autonomy level.

Soft Guardrails

Contextual boundaries that may relax over time. Require human approval initially but can be delegated later.

3 Dynamics & Evolution

Feedback Loops

How the pair learns and improves. Mechanisms for human corrections, agent suggestions, and mutual adaptation.

Capability Expansion

Anticipated new abilities that may emerge. How the pair will handle skills neither anticipated at the start.

Skill Drift Monitoring

How to detect when human skills atrophy or agent performance degrades. Maintaining healthy pair dynamics.

Evolution Metrics

How to measure relationship maturity. Indicators that the pair is ready for the next level of collaboration.

Example

HAP Plan
HAP PLAN: Financial Analyst + Research Agent
1. PAIR GOAL & SETUP
Shared Objective: Produce investment recommendations for the portfolio committee
Human Role: Strategic judgment, client relationships, final recommendations, accountability
Agent Role: Data gathering, quantitative analysis, report drafting, monitoring alerts
Shared KPIs: Recommendation accuracy, time-to-insight, client satisfaction scores
Tooling: Bloomberg terminal, internal CRM, research database, collaboration workspace
2. GATES & GUARDRAILS
Confidence Gate 1: After 50 reviews with <5% override rate, agent can auto-generate
standard company profiles without human review
Confidence Gate 2: After 6 months with positive feedback, agent can draft preliminary
recommendations (human approval still required before delivery)
Hard Guardrails:
- Never communicate directly with clients
- Never access personal client financial data
- Never make or suggest specific trade executions
Soft Guardrails:
- Flagging news items (can graduate to auto-summarizing)
- Scheduling research tasks (can graduate to auto-scheduling)
3. DYNAMICS & EVOLUTION
Feedback Loops:
- Weekly review of agent suggestions with explicit approval/rejection logging
- Quarterly analysis of override patterns to identify training gaps
- Agent proposes process improvements; human decides adoption
Capability Expansion:
- Anticipated: Sentiment analysis of earnings calls
- Anticipated: Cross-portfolio correlation detection
- Protocol: New capabilities require 30-day supervised trial
Skill Drift Monitoring:
- Track human's manual analysis frequency (should not drop to zero)
- Monthly "spot checks" where human validates agent work independently
- Alert if agent accuracy drops below 90% on previously mastered tasks
Evolution Metrics:
- Trust Score: % of tasks delegated vs. available for delegation
- Pair Velocity: Time-to-recommendation compared to baseline
- Maturity Level: Beginner → Developing → Established → Expert

Why This Works for Collaboration

This HAP Plan captures elements that neither User Stories nor Agent Stories can express: the evolving relationship between analyst and agent, the trust-building milestones that unlock autonomy, and the mechanisms that ensure the human remains capable and accountable even as the agent takes on more work. The pair is designed to grow together, with clear gates preventing premature automation and feedback loops ensuring continuous improvement.

Key Differentiators

Models the Relationship

Captures how trust builds over time, not just static task allocation.

Trust-Building Milestones

Defines concrete gates that unlock increased autonomy based on demonstrated performance.

Emergent Capabilities

Anticipates and manages abilities that arise from the collaboration itself.

Autonomy with Accountability

Balances agent independence with human responsibility and oversight.

Human Judgment Preserved

Ensures critical human capabilities don't atrophy as agent takes on more work.

Continuous Evolution

Built-in feedback loops and metrics for ongoing improvement.

Choosing the Right Artifact

Aspect User Stories Agent Stories HAP Plan
Focus Human goals Agent capabilities Pair relationship
Autonomy None Full Graduated
Evolution Static Capability-based Trust-based
Human Role Decision-maker Supervisor Partner
Best For Traditional apps Background automation Collaborative AI