Why I Reject the Equation Agent = LLM + Harness

1. Return to Cognitive Foundations: M+P+T as the Agent Soul

To discuss what an Agent truly is, we must revisit how the community’s understanding formed. AutoGPT was among the first projects to push the "LLM Agent" concept into practice, using a self‑loop where the model decomposes goals, calls tools, and iterates based on feedback. Lilian Weng’s widely‑cited paper LLM Powered Autonomous Agents distilled this practice into four components:

LLM (brain) + Planning + Memory + Tool Use . This M+P+T paradigm captured the core that gave early agents life: planning, memory, and tools expand the model’s capabilities, turning it from a text generator into an entity capable of long‑term reasoning and environmental interaction.

2. Re‑examining Harness: The Feedback Substrate

LangChain treats Harness as part of the Agent, but a more precise view is that Harness serves as the Agent’s feedback substrate, analogous to the environment in reinforcement learning. The harness does not replace the Agent’s reasoning; it provides high‑quality closed‑loop feedback.

Observation : When the Agent takes an action, the Harness runs the code in a sandbox and returns errors or results.

Signal : Through integration tests, linters, or validation gateways, the Harness tells the Agent whether its path is correct.

Thus the Harness is a mirror reflecting the Agent’s actions, while the interpretation of those signals and subsequent planning remain within the Agent’s cognitive architecture.

3. Harness Engineering Should Not Be Elevated to Definition

The author acknowledges the practical importance of Harness—current LLM planning is unstable, so engineering scaffolding is a reasonable short‑term strategy. However, conflating this engineering compromise with the definition of an Agent is problematic. Describing the present need for Harness is a pragmatic observation; stating Agent = LLM + Harness imposes a future‑locking ontological constraint.

If the community adopts this equation, engineers may default to adding more If‑Then logic in the Harness layer rather than enhancing the model’s autonomous reasoning, gradually narrowing the direction of Agent development.

4. Feedback Will Not Disappear: Co‑evolution of Agent and Harness

Some argue that as model capabilities grow, Harness will be absorbed and vanish. The author disagrees, likening it to a top programmer still needing an independent tester. Even a super‑intelligent model will require a deterministic environment interface to supply real‑world feedback and act as a safety fence.

The boundary between Agent and Harness shifts with model ability. Today, Harness may hard‑code error‑handling because the model cannot reliably classify errors. In the future, the model may internalize that logic, leaving Harness only to report what went wrong.

5. Beware the Degeneration Risk of "Agentic Workflow"

Andrew Ng promotes an "Agentic Workflow" where structured iteration lets GPT‑3.5 in Agent mode outperform zero‑shot GPT‑4. While commercially valuable now, treating Agent = LLM + Harness as truth can push development toward a deterministic workflow:

Agentic Workflow : Control resides in code (Harness); the path is a predefined If‑Then sequence—stable but autonomy‑limited.

Autonomous Agent : Control resides in the model, which dynamically generates task flows—exploratory and adaptable.

If Harness monopolizes decision branches and exception handling, we merely wrap the model in legacy automation scripts, a dimensionality reduction that stalls true Agent breakthroughs.

Conclusion

The author reiterates the rejection of the equation not because Harness is unimportant, but because it risks solidifying a temporary engineering compromise into the ontology of Agents. Harness is essential infrastructure, not a defining element. Progress should focus on refining feedback mechanisms so that Agents, powered by Memory, Planning, and Tools, can make complex autonomous decisions with minimal hard‑coded scaffolding.