dhaupin

One system, not a list of repos: deterministic guardrails, memory, and adversarial QA.

A vertically integrated, self-verifying software factory for AI-native software. Each tool is a layer of one idea: agent-written code needs the same discipline human-written code needs, and more of it, because the agent writes faster than a human reviews.

webappecomenterpriseagentorchestration

The system, in three layers

The family is one machine with three jobs. Each layer is a set of tools that refuse rather than warn.

Agent OS and memory

Persistent, auditable memory

A memory and runtime for AI agents, over MCP: markdown files in your own repo, read at session start and written before it ends, so git history is the memory lineage.

Fleet governance

One manifest, one spec, one factory

The family governs itself: a roster manifest renders the crosslinks, a spec holds the contracts, and a factory mints the next member.

Adversarial QA

Attack the code, audit what shipped

Attack the handlers in-process, watch the pixels with a committed visual memory, audit the baked output, and publish with no drift.

The one sentence

packstruct approves what the packs declare, crashtest grades what the code does, eyestruct explains what the pixels show, and poststruct probes what actually shipped, while govstruct governs who the members are and monoforge mints the next one.