Home · Getting Started · Model · Assessment Hub · Roadmap · References
Level 5 — Autonomous Delivery
License: CC BY-SA 4.0
https://creativecommons.org/licenses/by-sa/4.0/
Dark Factory
At this level, software production operates as a largely autonomous system.
Agents generate implementations, validate them against real-world behavior in digital-twin environments, and iterate until convergence — without human code writing or line-by-line review.
Confidence derives from a rigorous evaluation infrastructure rather than artifact inspection.
Harnesses, simulations, scenario-based validation, and real-world feedback loops collectively assure that outcomes meet intent.
Core Characteristics
- Spec → task graph → agent execution → validated outcome
- Fully automated build–test–revise cycles
- Continuous improvement within defined constraints
- Code treated as disposable implementation, not the primary asset
- Small human teams capable of maintaining very large systems
- Regeneration of subsystems is feasible as specifications evolve
Organizational Signals
Organizations at this stage often exhibit:
- Minimal human coordination for routine development work
- Industrial-grade auditability, traceability, and rollback capabilities
- Clearly defined domain boundaries where autonomy is permitted
- Explicit documentation of safe operating envelopes
- Outcome governance replacing implementation oversight
- Strategic focus on system-level risk management
Primary Bottleneck
Evaluation completeness and risk management
As autonomy increases, the central challenge becomes ensuring that evaluation mechanisms cover the full spectrum of realistic scenarios, failure modes, and unintended consequences.
Common Failure Mode
Dark Factory Mythology
Organizations claim full autonomy while significant hidden human labor continues to perform safety, verification, or corrective work.
Symptoms include:
- Undocumented manual interventions
- Reliance on expert operators to maintain stability
- Incomplete audit trails
- Overstated autonomy in external messaging
- Fragile systems that degrade without continuous human oversight
Critical Insight
Autonomy is domain-specific.
Different domains — due to technical complexity, regulatory requirements, safety considerations, and environmental variability — will achieve different levels of autonomy.
Most organizations will operate across multiple levels simultaneously:
- Legacy cores remain human-curated
- Peripheral or well-bounded systems become autonomous
- Migration occurs unevenly over time
Optimizing at Level 5
At this stage, the focus shifts from increasing autonomy to sustaining it safely.
Key ongoing concerns include:
- Governance of autonomous decision-making
- Long-term system evolution
- Ethical and regulatory compliance
- Monitoring for drift from intended behavior
-
Maintaining alignment between specifications and real-world outcomes
↑ Back to top · Home