Research

How AI-generated work survives contact with change.

We study how AI-native work can remain revisable and maintainable as goals, information, and decisions change. This research defines the mechanisms behind the ThruWire Pattern and the product systems we build with software companies.

Published papers

arXiv:2605.06365

From Agent Loops to Deterministic Graphs: Execution Lineage for Reproducible AI-Native Work

Josh Rosen and Seth Rosen · Submitted May 7, 2026

Introduces execution lineage: AI-native work represented as a directed acyclic graph of artifact-producing computations with explicit dependencies, stable intermediate boundaries, and identity-based replay.

Product implication

For a customer-facing product, execution lineage makes it possible to update the work affected by new information while proving what ran, what was reused, and what remained unchanged.

View on arXiv Read PDF

arXiv:2605.12087

Intermediate Artifacts as First-Class Citizens: A Data Model for Durable Intermediate Artifacts in Agentic Systems

Josh Rosen and Seth Rosen · Submitted May 12, 2026

Formalizes durable intermediate artifacts as typed, structured, addressable, versioned, dependency-aware work products that humans and agents can inspect, revise, supersede, and improve.

Product implication

For a customer-facing product, durable artifacts keep goals, evidence, judgments, plans, and decisions as maintained product data after the assistant finishes.

View on arXiv Read PDF

Research into product

The research informs all three parts of the pattern.

Data

Durable artifacts define the maintained goals, evidence, decisions, plans, and outcomes that become each customer’s working model.

Technology

Execution lineage gives AI assistants a reliable way to inspect, run, update, and selectively replay product-owned work.

UX

Explicit state and lineage make changes inspectable so customers can correct assumptions, approve decisions, and redirect the work.

Current agenda

What the product should know, maintain, and improve.

Our current work focuses on customer-specific working models, feedback and improvement, authority, change propagation, and systems organized around customer outcomes.

Apply the research to your product