Back to the newsroom

Op-ed

Agents fail not for lack of intelligence but for lack of constraints

Korea Deep LearningPublished

Agents fail not for lack of intelligence but for lack of constraints
In a January 2026 AI Times contribution, Korea Deep Learning pointed out that agents fail not because of intelligence but because of a lack of reproducibility and constraint design.

Kim Dong-hyun, head of Korea Deep Learning's Future AI Development Team, contributed "Agents fail not for lack of intelligence but for lack of constraints" to AI Times.

2025, a year that stayed at demos and pilots

The contribution starts with the report card on the 2025 AI agent boom. Compared with expectations, the results were mostly flashy demos and cautious pilots. Agents that worked well in text environments piled up errors on real work screens and in multistep procedures and wobbled in exceptional situations, earning the verdict of "automation too unsafe to leave in the field."

The problem is reproducibility, not intelligence

The author finds the heart of the problem in reproducibility. LLM agents behave probabilistically, so their output can differ even for the same input, and when documents or data are incomplete, hallucinations arise that plausibly fill in the gaps. That can be an advantage in content generation, but it is fatal in business work. The diagnosis is that what businesses want is not smarter answers but predictable execution, accountability and auditability.

"Businesses can only operate, and accountability and auditing only become possible, when the same input produces the same result."

— Kim Dong-hyun, Head of the Future AI Development Team, Korea Deep Learning

A hybrid approach and the value of judgment

The contribution cites the case of Salesforce, which saw the randomness of LLMs as a wall to reliability and shifted toward combining them with deterministic automation. Autonomy and the level of control should be designed differently depending on how critical the work is, and the key, it explains, is not to use LLMs less but to keep LLMs from acting arbitrarily.

It also argues that the place where revenue and risk reduction actually happen is not making card news or summarizing meeting minutes but the area of "judgment." The value of insurance document work lies not in summarizing but in sorting whether requirements are met into pass or fail. A stable structure, it says, handles what can be defined by rules, such as missing required fields, signatures and dates, and the suitability of supporting documents, with rules, schemas and triggers, while for ambiguous exceptions the LLM puts forward evidence and candidates together and then passes them through a validator.

"This way, hallucination can be handled not as a 'wish to eliminate it' but as a design that 'stops' it from turning into an incident."

Three constraints that decide success or failure

The author sets out three kinds of constraint design that business agents need. Output restriction fixes the result format and the allowed range to reduce free-form writing. Enforced verification sets up rule checks and cross-validation and makes stopping or escalating to a higher level the default when a check fails. Least privilege narrows the tools that can be called and the execution permissions, and makes everything traceable through logs and replay. The conclusion is that competition in the age of agents depends not on bigger models but on how precisely a constrained execution space is designed, and that document-based work is the best area for defining such constraints.

The full text appeared in AI Times.

About Korea Deep Learning

Korea Deep Learning (KDL), founded in 2019, is an enterprise AI agent company. Built on KDL Frontier, its own vision-language model (VLM) specialized for documents, KDL makes AI agents that take on and complete the document work of businesses and public institutions. Its main products are DEEP OCR, DEEP Parser, DEEP Agent and DEEP Ops.

Source