When the Agent's Own Output Becomes Part of the System Running It

On a framework for governing runtime self-modification, and the substrate it assumes already exists.

On May 26, 2026, Mariano Garralda-Barrio published Governed Evolution of Agent Runtimes through Executable Operational Cognition on arXiv, introducing a mechanism called HarnessMutation for managing a problem that most agentic AI architectures have not yet named explicitly: what happens when an agent's own generated output stops being a disposable answer and starts becoming a persistent part of the system that produced it.

The paper opens from a premise that is easy to state and consequential to accept: recent advances in agentic systems increasingly treat code as an executable operational substrate rather than as a disposable output artifact. Prior work, which the paper cites as "Code as Agent Harness," frames validated agent-generated artifacts as runtime entities that can be created, executed, revised, persisted, and reused within long-running cognitive loops. Once an agent's output is not just an answer but a piece of code that will run again, and again, and be modified by the agent itself over time, the governance question changes shape entirely. The paper's diagnosis is blunt: the governance, lifecycle management, and operational evolution of such artifacts remain under-specified.

The Mechanism

HarnessMutation is introduced as a governed mechanism for lifecycle-aware runtime adaptation operating under explicit validation, traceability, evaluation, and rollback constraints. Each of those four constraints does distinct work. Validation determines whether a proposed change to the runtime is acceptable before it takes effect. Traceability means the change is recorded in a way that allows later reconstruction of what changed and why. Evaluation means the change's effect on system behavior is measured, not assumed. Rollback means the change can be reversed if evaluation reveals a problem after the fact.

The paper is careful to draw a boundary around what it is proposing versus what it is explicitly rejecting: rather than treating runtime adaptation as unrestricted self-modification, the proposed framework models evolution as a bounded and observable process over persistent operational memory. That distinction, bounded and observable versus unrestricted, is the paper's central contribution. It is not arguing that agents should be prevented from modifying their own operational substrate. It is arguing that if they do, the modification needs to happen inside a framework where each of the four constraints applies at every step, not as an occasional audit layered on afterward.

Formalizing the Artifact as Infrastructure

The paper's broader theoretical move is to formalize agent-generated artifacts as persistent runtime capabilities that progressively become part of the operational substrate rather than transient intermediate outputs. This formalization matters because it changes what kind of thing needs to be governed. A transient output can be discarded if it turns out to be wrong. A persistent runtime capability that has already been incorporated into the operational substrate cannot simply be discarded without addressing everything downstream that came to depend on it. The paper is naming a category of artifact that most current governance frameworks, built around reviewing outputs before they are used, are not structured to handle, because those frameworks assume the artifact under review is not itself going to become infrastructure.

Convergence With a Parallel Paper

One day earlier, on May 25, 2026, Shangding Gu's Scaling the Harness in Agentic AI independently argued that the harness, the structured execution layer surrounding a foundation model, is the field's next major bottleneck, and proposed measuring properties like trajectory quality, memory hygiene, and verification cost at that layer. Two papers, on consecutive days, arriving from different starting premises, both treat the harness not as an implementation detail but as a first-class object requiring its own governance architecture. Garralda-Barrio's contribution is specifically about the case where the harness itself changes at runtime, driven by the agent's own outputs, a case Gu's paper frames primarily in terms of static architectural components.

What the Framework Assumes Without Establishing

HarnessMutation is described as operationalizable over modern agent runtimes and governance-oriented orchestration systems, providing a conceptual foundation for adaptive infrastructures whose evolution remains explicit, auditable, and constrained. The paper is explicit that this is a conceptual foundation, not a deployed system. It does not claim that any production agent runtime today implements the four constraints, validation, traceability, evaluation, and rollback, as a mandatory gate on self-modification. It proposes the shape the gate should take. Whether any deployed system actually enforces that shape, rather than allowing agent-generated artifacts to accumulate into the operational substrate without the governed checkpoints the paper describes, is a separate and unanswered question.

Open Questions

• How many currently deployed agent systems that persist agent-generated code as runtime capability actually implement validation, traceability, evaluation, and rollback as mandatory gates, rather than optional review steps?
• At what point does an agent-generated artifact cross from "transient intermediate output" to "part of the operational substrate," and who makes that determination in a live system?
• What standard applies when a rollback is triggered after an agent-generated artifact has already been incorporated into other downstream processes that depend on it?
• How does traceability survive when the artifact being traced was itself produced by a stochastic process that cannot be exactly reproduced on demand?
• Who audits the auditor: what verifies that the evaluation step within HarnessMutation is itself functioning correctly, rather than approving changes it should have blocked?
• Given that two independent papers converged on harness governance within a single day of each other, what does that convergence suggest about how far behind deployed practice sits relative to the research frontier?

The loop closed around an oversight function that was never instrumented.