The model won't devour everything: three primitives of agent systems

Software researcher Bowen Liu published the v1.0 specification of the Runtime-Workspace-Framework in October 2026, defining the minimal three primitives required for autonomous agent systems. Derived from five streams of empirical evidence and four regularities, the work argues that frontier models will not subsume every agent component, but will progressively replace acceptance-executable, non-physical parts across an observation window running through late 2028. The release is detailed in github.com's report along with formal propositions, verified 2026 arXiv citations, and operational acceptance criteria for eight testable architectural predictions.

This analysis details the concrete primitives, structural corollaries, and evaluation criteria governing agent infrastructure. It provides engineering teams, infrastructure architects, and system evaluators with precise definitions for where runtime frameworks thin out and how persistent state survives base-model iteration.

Model won't devour everything: what actually changed

The framework establishes that autonomous systems reduce to three distinct primitives: Runtime, Workspace, and External Knowledge Sources. Under the closed-loop structure law, the runtime serves as the execution process layer handling load operations, system calls, scheduling, and interrupts. The model itself functions as an ephemeral callable function, while the runtime functions as the enclosing operating system process. The workspace represents the persistent state layer, serving as the sole architectural structure that survives across working sessions and base-model transitions.

Alongside runtime and workspace, external knowledge sources introduce what the author terms the world term under the collapse law. These sources represent real external signals collected on their own cadence that enter the workspace strictly through an explicit loading path. By separating the dynamic execution process from persistent workspace state and external knowledge inputs, the framework eliminates ambiguity about which operational components can be automated by foundation models and which components require external architectural enforcement.

Model won't devour everything: how it works

System evolution under this framework produces four structural corollaries. First, the runtime layer thins over time as base-model iteration absorbs judgment functions like loading strategy, scheduling, and request routing. However, the runtime's enforcement responsibilities—including sandbox boundaries, permission controls, and credential management—persist indefinitely because models of arbitrary capability still require strict operating constraints. The runtime ultimately converges toward the minimal footprint of an operating system kernel.

Second, fine-tuning and model distillation shift from being a standard pipeline step into a conditional branch. Distillation triggers only when four conjunctive conditions hold simultaneously: high call frequency, stable distribution, automatically verifiable outputs, and sustained holdout superiority over the latest base model paired with workspace state. Third, an asset hierarchy ranks models lowest, followed by runtimes, with workspaces ranked highest. Workspace preservation operations are limited to backups, lineage tracking, and evaluation-set maintenance. Fourth, human handover depth follows acceptance executability, positioning humans strictly at specification and acceptance adjudication.

Model won't devour everything: why it matters now

The architecture directly addresses current fragmentation across agent orchestration software. Empirical measurements recorded a 48-point performance spread across different runtime implementations on a single grade-B benchmark source. The framework predicts this benchmark spread will narrow by more than half across two or more independent reproducible measurements as context loading shifts from explicit runtime configuration files to model-driven retrieval within two base-model generations.

Furthermore, graph-coded orchestration frameworks face structural obsolescence. The paper outlines operational acceptance criteria showing that graph-based agent orchestration tooling will marginalize within 12 to 24 months through vendor pivots, acquisitions, or documented declines in deployments and open-source activity. As base models absorb execution graph logic natively, runtime layers that focus solely on static prompt chains or rigid state graphs lose technical relevance, shifting value toward evaluation-set design and persistent state preservation.

Model won't devour everything: who is affected

Engineering roles, developer tooling startups, and infrastructure teams experience immediate shifts based on how acceptance criteria are adjudicated. Where task acceptance is not computationally executable, the author warns that human requirement-writing degrades into making unverified wishes. Consequently, human labor concentrates at the boundaries of task specification and acceptance adjudication, while roles centered on easily automatable verification criteria experience steeper volume declines.

Specialized workspace management is also emerging as its own independent product category. Tooling teams previously building static execution wrappers are redirected toward building dedicated services for workspace backup, data lineage, and evaluation-set curation. Meanwhile, machine learning teams relying on default distillation workflows must restructure their pipelines to treat fine-tuning as an exception, testing candidate distilled models against baseline configurations of the latest base foundation model augmented with persistent workspace context.

Model won't devour everything: what to watch

The framework establishes eight testable predictions with expiry dates through late 2028, each tied to an operational acceptance criterion. Prediction three requires sandbox, permission, and credential layers to remain backward-compatible and never be natively replaced by model weights. Prediction six states that Level 5 autonomous systems—defined as end-to-end autonomous successor-model training—will yield zero reproducible public implementations across code and data within the late-2028 observation window.

The author also leaves three open problems explicitly documented for community inspection: loop dynamics, cross-repository migration, and the formalization of curation paths. These unresolved mechanisms represent incomplete boundaries of the v1.0 specification rather than counterexamples to the core primitives. Progress will be scored strictly at prediction expiration against the published criteria, tracking whether evaluation-set design appears in mainstream job postings and whether flagship runtimes make model-driven retrieval their default path.