Without changing the model, simply adjusting the Agent’s execution framework raised task accuracy from 52.8% to 66.5%. We notice that competition in AI Agents is shifting from model capability toward the engineering collaboration between models and business systems.
What this is
An Agent can be understood as an AI system that calls tools and executes tasks in multiple steps; a Harness is the execution framework wrapped around the model, responsible for context management, permission decisions, failure retries, and workflow orchestration. The article summarizes this as “Agent = LLM + context + tools.”
Current Agents have two obvious gaps. First, they are not good at real-time interaction and typically respond only after an entire reasoning sequence has finished. Second, they are not good at continuously accumulating experience: after a task ends, what remains is usually temporary context rather than reusable methods and knowledge.
Models are now filling part of this gap. For example, 小米 MiMo generates more than 1,000 tokens per second for a 1-trillion-parameter model on a single eight-GPU node; Taalas HC1 embeds an 8-billion-parameter model directly into a chip, delivering roughly 17,000 tokens per second and a response time below 100 milliseconds, but updating the model requires a new chip fabrication run.
Industry view
Our judgment is that models will progressively absorb general-purpose framework code, but they will not eliminate the Harness—the external system responsible for orchestration, constraints, and fallback handling. Every time a model internalizes a capability, the corresponding patch may lose its value; yet permission rules, time-sensitive requirements, and exception handling in real businesses will remain outside the system for the long term.
This path also carries risks. Enterprises may bind critical workflows too early to rapidly updated models or frameworks, bearing migration, maintenance, and revalidation costs. At the same time, recording task trajectories and handling enterprise data create privacy and compliance responsibilities in their own right. A safer approach is to keep general capabilities continuously modularized while retaining data permissions, human approvals, and performance evaluation on the side enterprises control.
The article also highlights an optimistic premise that warrants caution: feedback from real-world scenarios must be able to flow legally and compliantly back into training. Otherwise, Agents can only generate logs, not close the capability loop. The long-term advantage of application companies may therefore lie not in the framework itself, but in exclusive data, stable channels, customer trust, and business scenarios that require collaboration between people and Agents.
Impact on regular people
For enterprise IT: procurement priorities will shift from “which models are integrated” to permission isolation, process audit trails, failure recovery, and performance evaluation. The maintenance cost of existing Agent projects will not automatically fall simply because models are getting faster.
For individual work: tools that interrupt, ask follow-up questions, and proactively advance tasks will become closer to real-time assistants, but enterprise rules will still limit what they can read and execute. Truly valuable experience must be organized into reusable knowledge rather than left to the expectation that tools will automatically remember everything.
For the consumer market: response latency will continue to fall, making voice and visual assistants more natural. Yet products that continuously understand personal preferences without leaking data will still depend on whether enterprises earn trust and whether users are allowed to control memory.