Skip to content

Governed Evolution of Agent Runtimes through Executable Operational Cognition

🕒 Published (v1): 2026-05-26 17:36 UTC · Source: Arxiv · link

Why this paper was selected

Governed harness evolution via executable operational cognition artifacts

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This systems vision paper proposes a governance framework for how agent-generated executable artifacts (prompts, evaluators, workflows, skills) should evolve from transient task outputs into persistent, lifecycle-managed runtime capabilities. It introduces HarnessMutation as a bounded, auditable transformation mechanism over harness configurations, and a Knowledge-Grounded Runtime Graph to track lineage and dependency among evolving artifacts. The core claim is that runtime evolution must be observable, reversible, and governance-constrained rather than unconstrained self-modification.

Problem

Prior work (notably "Code as Agent Harness") establishes that agents produce executable artifacts inside task loops, but leaves their governance, lifecycle management, and operational evolution underspecified. Without explicit lifecycle control, accumulating agent-generated capabilities risk capability drift, silent regressions, evaluation contamination, and non-auditable operational behavior. The specific gap is: once agent-initiated artifacts persist and influence future behavior, how should their promotion, mutation, deprecation, and safety be formally modeled?

Method

The framework rests on three coupled constructs:

  1. Harness formalization: A harness \(H = \{P, T, E, M, G, O, K\}\) where \(O\) (executable operational artifacts) and \(K\) (structured operational knowledge) are new first-class components alongside prompts, tools, evaluators, memory, and governance constraints. The optimal harness at iteration \(t\) is: $\(h^* = \arg\max_{h_i \in C_t} F(h_i) = \alpha Q(h_i) + \beta R(h_i) + \gamma V(h_i) + \delta U(h_i) - \lambda C(h_i)\)$ where \(U(h_i)\) is a reuse-value term absent from prior task-level formulations.

  2. HarnessMutation \(\mu: h_i \to h_i'\): a governed transformation carrying a bounded change contract specifying the component modified, targeted failure mode, expected improvement, preserved invariants, a falsifying evaluator, and rollback conditions. Mutations remain candidates subject to lifecycle review rather than being directly adopted.

  3. Capability Lifecycle \(L = \{\text{experimental}, \text{validated}, \text{trusted}, \text{canonical}, \text{deprecated}\}\): artifacts transition through states via evidence-driven governance. A Knowledge-Grounded Runtime Graph \(G_t = (V_t, E_t)\) with typed edges \(R = \{\text{depends\_on, generated\_by, validated\_by, improves, supersedes, mutated\_from, composed\_with, fails\_under}\}\) tracks lineage and operational relations. Each node carries \(\varphi(v_i) = (c_i, q_i, \tau_i, \ell_i)\) with quality score \(q_i = \omega_p p_i + \omega_r r_i + \omega_s s_i + \omega_u u_i - \omega_\rho \rho_i\).

Key Contributions

  • Executable operational cognition: formal framing in which agent-generated artifacts become persistent runtime capabilities rather than transient outputs.
  • HarnessMutation: a lifecycle-aware, rollback-capable transformation primitive over prompts, workflows, evaluators, routing policies, and skills.
  • Capability lifecycle model: five-state progression (experimental → validated → trusted → canonical → deprecated) separating generation from adoption.
  • Knowledge-Grounded Runtime Graph: structured representation of lineage, dependency, validation, mutation history, and composition relations among evolving artifacts.
  • Governance-oriented multi-agent architecture: specialization of agents into generation, validation, evaluation, governance, and execution roles over a shared governed runtime kernel.

Results

This is a conceptual/architectural systems paper; no empirical experiments or benchmark numbers are reported. The paper cites external empirical work (Voyager, SkillOpt) as precursor evidence for specific sub-claims (executable skills as reusable operational memory; bounded textual-edit mutation with held-out validation gates) but does not produce new measurements.

Limitations

  • Purely conceptual: no implementation, prototype, or empirical validation of the proposed framework.
  • HarnessMutation governance contracts and lifecycle transition criteria are described schematically; concrete algorithms for deciding promotion or deprecation are left to future work.
  • The Knowledge-Grounded Runtime Graph is presented architecturally; scalability, graph maintenance cost, and query mechanisms are unaddressed.
  • The multi-dimensional objective \(F(h_i)\) requires specifying \(\alpha, \beta, \gamma, \delta, \lambda\) and component quality scores (\(Q, R, V, U, C\)) — no guidance on eliciting or learning these weights is given.
  • Restricted to synchronous lifecycle governance; dynamic or adversarial artifact injection scenarios are not analyzed.

Relevance to Harnesses / Meta-Harnesses

This paper is a direct theoretical extension of the "Code as Agent Harness" framing and is among the first to treat the harness itself as an evolvable, governed object rather than fixed infrastructure. For researchers building or studying meta-harnesses, it provides a formal vocabulary for the problem of harness self-modification: how a running harness should validate, version, promote, and roll back its own operational components. The HarnessMutation primitive and capability lifecycle model are directly applicable to meta-harness designs that must accumulate and curate agent-generated workflows, evaluators, and routing policies over time. The runtime graph's typed lineage edges also address a persistent gap in existing harness frameworks — the absence of structured provenance for harness components that have been mutated or composed by sub-agents.