Will Stronger Models Render Harnesses Obsolete? (Part 2)

Model‑Driven Removal of Harness Components

In November 2025 Anthropic released its first technical paper on Harness, which at the time relied on two core mechanisms: Context Reset , which cleared the session context at the start of each new conversation, and Sprint Contract , which forced the Generator and Evaluator to agree on acceptance criteria before a work unit began. After the release of a new Claude version, Anthropic eliminated both mechanisms because the upgraded model’s built‑in context management made the reset unnecessary and the model could autonomously pace its execution without an explicit sprint contract.

The LangChain team expressed a similar view: as model abilities continue to improve, previously hand‑crafted inductive biases become redundant, and the framework‑level scaffolding can be torn down.

Capability Migration: From External to Internal

Historically, Agent frameworks such as LangChain helped developers assemble tool calls, multiple Agent roles, and multi‑step reasoning pipelines. Today, many of these functions are handled natively by stronger models: the model decides when to invoke a tool, when to delegate to a sub‑Agent, and when a task is complete, eliminating the need for explicit orchestration instructions.

However, a small set of capabilities remain outside the model’s reach because they are fundamentally infrastructure concerns rather than inference problems. Persistence requires external storage, cost control needs awareness of organizational budget limits, and audit logs must satisfy regulatory compliance—none of which the model can infer on its own.

Consequently, over the past three years AI systems have shown a consistent migration of ability from the “external” layer (system prompts, Agent loops, tool‑management logic) to the “internal” layer (model weights, activation values, forward propagation).

Tool‑calling, once a special framework feature, is now a native model capability.

Multi‑step reasoning, formerly guided by explicit Chain‑of‑Thought prompts, is now performed automatically by strong models.

Context compression, previously implemented by external summarisation logic, is now decided by the model itself.

Boundaries of Current Harnesses

Three problem categories expose the limits of existing Harnesses:

Diagnostic errors : the Agent correctly executes a task but follows the wrong direction; Harness can detect format violations but cannot recognise that the entire logical path is mistaken.

Over‑engineering : the Agent adds unnecessary abstraction layers or overly complex solutions, requiring deep business‑context understanding that Harness cannot provide.

Semantic misinterpretation : the Agent follows the literal instruction but misreads the true intent, e.g., splitting a function into five clear sub‑functions that break all callers.

These issues all demand a deep understanding of “what the correct result is,” not merely a formal validation of output format. Current Harness toolchains excel at the latter and struggle with the former, meaning the ultimate remedy lies in further model improvements.

Three Types of Harness That Won’t Be Eaten

Based on the analysis of computing history and current AI‑engineering practice, three Harness categories are unlikely to disappear as models get stronger:

Persistence and External State Management

Models lack cross‑session memory; therefore, external storage reads/writes, database transactions, and filesystem operations will always require Harness mediation.

Compliance Auditing and Permission Gating

Enterprise agents that can run code, send emails, or modify databases must have explicit permission boundaries and audit trails to satisfy legal and regulatory requirements—constraints that are institutional rather than technical.

Cost Control and Observability

Models cannot perceive budget limits or token‑usage caps; external monitoring and control layers are needed to enforce constraints such as “no single task exceeds $40.”

Enterprise Paradox: Autonomy vs. Compliance

Granting Agents greater autonomy can unlock the model’s full potential, but it also raises the risk of catastrophic failures (e.g., accidental data deletion or unauthorized API calls). Unless the model provider assumes liability, enterprises must retain human‑in‑the‑loop checks, permission gates, and pre‑execution confirmations for critical operations.

Co‑evolution of Harness and Model

Debugging and optimizing Harnesses yields valuable execution‑trajectory data: which tool‑call sequences succeed, where agents stall, and which contexts cause format violations. This data can be fed back into model training, making the Harness both a usage environment and a source of training signals.

When a Harness component encodes an assumption about a model’s current limitation, that assumption must be re‑validated after each model upgrade; obsolete components should be removed rather than left as technical debt.

Engineering Recommendations

Delete‑Oriented Programming

Before adding any Harness component, record the specific model deficiency it compensates for and the capability threshold at which the component can be removed. For example, “When the model’s sub‑task omission rate on X‑type tasks falls below 5 %, delete the task‑decomposition verification component.”

Make Observability Core

Observability should be built in from the start, collecting metrics such as tool‑call failure rates, agent stall points, and added latency. Without this data, Harness improvements rely on intuition rather than evidence.

Treat Trust and Compliance as Infrastructure

Audit logs, permission controls, and irreversible‑action confirmations must remain regardless of model strength because they address trust, accountability, and regulatory compliance—not model capability.

Conclusion

Just as past technological shifts have shown, models will continue to “eat” the layers of Harness that compensate for their current shortcomings. However, the need to steer the horse—managing persistence, compliance, cost, and observability—remains essential, even as the harness becomes lighter and more sophisticated.