The Bottleneck Moved and the Metrics Have Not Caught Up

On a paper arguing that agentic AI's next constraint is not the model but the structure built around it.

On May 25, 2026, Shangding Gu published Scaling the Harness in Agentic AI on arXiv, arguing that the field's next major bottleneck is system scaling, not model scaling. The paper's framing is direct: the design of auditable, persistent, modular, and verifiable architecture around a foundation model, what the paper calls the harness, is now the constraint that determines whether an agent behaves reliably over a long-horizon task, not the raw capability of the underlying model.

The paper's core claim is that current evaluation practice has not caught up to this shift. Standard benchmarks remain largely model-centric, often reducing agents to final-task success while treating memory, retrieval, tool use, orchestration, verification, and governance as secondary implementation details. That framing, the paper argues, is increasingly inadequate, because agent performance actually emerges from the interaction among the foundation model, memory substrate, context constructor, skill-routing layer, orchestration loop, and verification-and-governance layer. Those six components, taken together, are what the paper defines as the harness: the structure that translates model capability into long-horizon agent behavior.

Naming Verification as a Cost, Not an Afterthought

The paper's most consequential methodological move is to propose harness-level benchmarks that measure properties current evaluation does not track: trajectory quality, memory hygiene, context efficiency, communication fidelity, verification cost, and safe evolution over time. Naming verification cost as a distinct, measurable metric is a departure from treating verification as a binary gate, present or absent, rather than a resource that scales with system complexity and that trades off against throughput, latency, and deployment cost.

This is a different framing than most governance conversations use. Governance discussions tend to treat verification as a compliance checkbox: does the system have a human-in-the-loop step, yes or no. Gu's paper treats verification cost as a first-class variable in the same category as memory hygiene or context efficiency, something with a measurable magnitude that a system architect has to budget for, not a binary condition satisfied once and then assumed to hold.

Three Named Bottlenecks

The paper structures its argument around three specific bottlenecks inside the harness: context governance, trustworthy memory, and dynamic skill routing, together with the orchestration and governance mechanisms that coordinate and constrain them. Each of these is a place where an agent's long-horizon behavior can silently diverge from its intended behavior without a single-turn evaluation ever catching it. Context governance determines what information the agent carries forward from one step to the next. Trustworthy memory determines what the agent believes it knows and how that belief was justified. Dynamic skill routing determines which capability the agent invokes for a given subtask. None of these three properties shows up in a benchmark that measures only whether the agent's final output matched an expected answer.

A Reference Implementation, Not a Standard

To make the argument concrete, the paper introduces CheetahClaws, a Python-native reference harness, comparing it against Claude Code and OpenClaw. The comparison is offered as an illustration of what a harness that treats these components as first-class objects looks like in practice, not as a proposed industry standard. The paper's stated main claim is unambiguous: future progress in agentic AI will depend as much on system design as on stronger foundation models.

Where This Lands in the Broader Convergence

Gu's paper did not arrive in isolation. One day later, on May 26, 2026, a separate paper by Mariano Garralda-Barrio introduced HarnessMutation, a governed mechanism for lifecycle-aware runtime adaptation of agent-generated artifacts, independently treating the harness as the substrate where governed evolution has to happen. Two independent papers, published on consecutive days, converging on the harness as a first-class object of design and governance, is the kind of convergence signal that suggests a structural gap in the field rather than a single author's idiosyncratic framing.

The Frontier Model Forum's own June 2026 issue brief reached a parallel conclusion from the security side, describing the harness as the layer where "agent actions pass through" and therefore "a natural control point for monitoring, verification, logging, and enforcement." Three separate efforts, an academic systems paper, an academic governance paper, and an industry consortium security brief, arrived at the same architectural focal point within roughly a month of each other, using different vocabularies for what amounts to the same claim: the harness is where the unsolved problem lives.

What Naming the Bottleneck Does Not Resolve

Naming verification cost as a metric is not the same as establishing what magnitude of that cost is acceptable, or who bears it. The paper proposes a research agenda for harness-level benchmarks; it does not claim those benchmarks yet exist in a form deployable across the industry, nor does it specify who would run them, at what cadence, or with what authority to act on the results. A metric that exists in a single research paper's proposed agenda is not yet an instrument any deployer is accountable to.

Open Questions

• What magnitude of verification cost should be considered acceptable for a given deployment context, and who determines that threshold?
• At what point does a harness-level benchmark move from a research proposal to an instrument deployers are actually measured against?
• How does an organization verify that its context governance, memory, and skill-routing layers are behaving as intended, absent the harness-level benchmarks the paper proposes but does not yet operationalize?
• If three independent efforts converged on the harness as the critical substrate within a single month, what does that convergence imply about the maturity of the tooling actually deployed today?
• Who is responsible for auditing a reference harness like CheetahClaws against the six-component framework the paper defines, once such harnesses are adopted outside research settings?
• Does treating verification as a cost rather than a binary gate change the incentive to skip it under deployment pressure, and if so, in which direction?

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