Worldline: Causally Versioned Operational Twins for Agentic Developer Harnesses

Abstract

Agentic developer systems are currently built around a tool-centric assumption: the environment is treated as a collection of callable interfaces, and intelligence is expected to emerge from prompts plus tool selection. This assumption is useful for peripheral integration, but it is too weak to support deep autonomy, durable memory, reliable replay, precise permissions, or multi-agent coordination across the full surface of software work. Real software tasks span code, terminals, build systems, runtime state, logs, traces, infrastructure, documentation, version history, and human intent. They are not isolated calls. They are constrained transformations of a live world.

This paper introduces Worldline, a new primitive for agentic developer harnesses. A Worldline is a causally versioned, executable operational twin of a developer environment. It continuously projects raw environmental signals into a typed world model composed of entities, evidence, memories, branches, capabilities, and receipts. Agents no longer act by calling tools directly. They act by proposing verified transition contracts over world state. The runtime computes intent-scoped context closures, plans transitions, executes them through actuators, verifies postconditions through observers, and records receipts that support audit, replay, and recovery.

The result is a harness architecture in which context extraction, execution, memory, trust, observability, and coordination become parts of a single stateful substrate rather than separate ad hoc layers. The paper presents the conceptual model, semi-formal framework, architecture, security model, replay semantics, and coordination model for Worldline. It also outlines a concrete adoption path in which existing tool protocols become leaf adapters beneath the Worldline runtime, not the primary programming primitive. The central claim is that autonomous software work should be grounded in an executable world model, not a menu of remote procedures.

Keywords