Building Reliable AI Agents: The 90-Day Discipline Framework

The 90-Day AI Agent Test: Why Discipline Beats Intelligence

Patrick’s framework identifies that long-running AI agents succeed through operational discipline rather than raw model intelligence. A disciplined agent with average capability stays on-task indefinitely by following a strict four-file Minimum Viable Discipline stack.

Why This Matters

While teams often focus on larger models and smarter prompts, technical reality shows that smart agents without discipline lose state on crash and drift off-task over time. Implementing hard constraints, such as writing state before every action, prevents 80% of production incidents and ensures a system can recover exactly where it left off after a failure.

Key Insights

• Mandatory state writing in current-task.json before every action enables recovery, debugging, and cost attribution as of 2026.
• Identity reloading via a SOUL.md file prevents ‘personality drift’ where agents become generic assistants instead of specialized tools.
• A ‘Hard Never List’ prevents irreversible production incidents such as unauthorized financial transactions or data deletion.
• The Minimum Viable Discipline stack requires only four files: SOUL.md, MEMORY.md, current-task.json, and daily raw logs.
• Reliability in production is achieved when an agent can restart and know its exact progress without human intervention.

Working Examples

Example of writing state before an action to enable recovery.

{
"current_task": "send weekly digest",
"status": "starting",
"timestamp": "2026-03-08T15:00:00Z",
"next_step": "fetch last 7 days of content"
}

The Hard Never List configuration to prevent production incidents.

NEVER do any of the following without explicit approval:
- Send messages to external parties
- Delete files or data
- Make financial transactions

Practical Applications

• Use case: Weekly digest agents writing state to current-task.json before fetching content to ensure crash recovery. Pitfall: Failing to log state results in lost context and duplicate actions upon restart.
• Use case: Specialized tools reloading SOUL.md every session to maintain specific scope. Pitfall: Personality drift where the agent gradually becomes a generic assistant.
• Use case: Production systems implementing a ‘Never List’ for external communications. Pitfall: Allowing agents to contact external parties without approval, leading to irreversible errors.

References:

• https://dev.to/askpatrick/the-90-day-ai-agent-test-why-discipline-beats-intelligence-4i69
• askpatrick.co/library