Your Software Factory Needs a Bill of Materials | Daniel Meppiel
Summary
As software development shifts toward 'skills' (prose/natural language) in AI-native factories, traditional versioning is insufficient. Daniel Meppiel argues that enterprises require a new 'fabric of trust' built around Agent Contracts, Agent Package Manager (APM), and signed receipts. This system provides traceability, provenance (via SLSA), and verifiable evidence, enabling automated, high-trust CI/CD gates and reducing reliance on human review.
Key takeaways
-
The Shift from Code to Skills
1:53
The industry is moving toward writing 'prose' or 'skills' rather than traditional code. These skills are organized into modules that form an automated 'factory,' which is a new programming paradigm.
-
Versioning the Factory, Not Just the Code
5:40
A factory must be versioned. The smallest unit of work is proposed as an 'Agent Contract'—a standardized definition of the agent's role, dependencies, and budget, which can be versioned and composed.
-
The Role of APM and ABOM
24:09
The Agent Package Manager (APM) functions like npm, managing the dependencies (skills, hooks, MCPs) required for a factory. This process generates an Agent Bill of Materials (ABOM), providing traceability.
-
Trust through Signed Receipts
36:59
A factory run generates a signed receipt. This receipt cryptographically chains the output (diffs) to the evidence (checks and ABOM), allowing CI/CD gates to validate the integrity and provenance of the change.
Technical details
-
Agent Contracts and Composition
2838s
An Agent Contract defines the assignment (input/output), the role, and the dependencies. A factory is viewed as a graph of these contracts (e.g., planning, specification, build, review) that chain together to perform a task.
-
Trust Fabric Components
2056s
The trust fabric relies on four elements: Verification (does the code meet intent?), Compliance (does it follow standards?), Traceability (can we trace back to the skill/factory?), and Security (is the factory itself secure?).
-
Security vs. Guidance
1744s
Skills are inherently 'guidance' (non-deterministic behavior), while security controls are 'deterministic' (runtime governance, sandboxes, permissions). These controls must be separate from the skills.
-
Verification and Deterministic Checks
2321s
Checks defined in contracts must be deterministic (e.g., linting, format checks). The goal is to shift testing left, providing cryptographic evidence that checks ran and passed, allowing CI/CD to skip redundant runs.
Mentioned resources
- APM (Agent Package Manager)
- SLSA and in-toto
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.