Uncovering the Realities of Enterprise Agent Deployment: Protocols, Costs, and Evolution

Topic 1: Overlooked Shortcomings When Agents Enter Business

Cost‑focused director Mao Zhuo stresses that an agent must deliver clear business value; otherwise massive token consumption makes the project unsustainable. He cites a calculation from his company (tens of thousands of employees) showing that widespread agent use could cost tens of millions of tokens annually, prompting a ROI analysis.

Technical lead Kuan Yuping points out that individual productivity gains (60‑80%) collapse to about 20% at the team level because of unstable context, lack of traceability, and ungoverned results. He argues that enterprises need a full‑stack, process‑oriented agent ecosystem rather than a better chat model.

Researcher Gu Yongfeng adds that agents lack inter‑connection; essential capabilities such as discovery, collaboration, trust, and reputation remain scattered across prompts, logs, and private APIs.

Topic 2: Foundation Protocol vs. Spec‑First – How Two Foundations Align

Gu clarifies that the protocol layer addresses how agents communicate and share information, likening it to a manager juggling many contacts on WeChat. Kuan agrees and defines the protocol layer as the long‑term infrastructure, while the process layer handles immediate workflow integration.

Mao illustrates a future where a single super‑agent acts as an operating‑system‑level service, making a standard protocol indispensable.

Topic 3: Protocol Layer vs. Tool Layer – Which Is Easier for Business Adoption?

Gu argues the tool layer wins because, like HTTP or TCP/IP, developers prefer ready‑made SDKs over low‑level handshakes.

Kuan notes that tools provide quick wins for early teams, but without a unified protocol, scaling leads to maintenance nightmares.

Mao emphasizes that business stakeholders care only about concrete outcomes (e.g., accurate data queries), so usability must trump technical elegance.

Topic 4: Token Cost Runaway – Technical Issue or Management Issue?

Mao says both factors matter. Technically, different IDEs and CLI tools vary in token efficiency; managerially, an individual's communication style can amplify token waste.

Kuan identifies three sources of token waste: repetitive injection of unnecessary context, unstructured knowledge layering, and lack of reusable process assets. He likens large models to a “seven‑second memory fish” that is also a “Swiss‑army‑knife”.

Mao shares an architectural tip: isolate each agent’s work to a dedicated microservice repository to constrain its understanding and code changes.

Topic 5: Cost Reduction vs. Efficiency Gains – Conflict or Unity?

Mao reports a monthly cost drop of 88% and a 3‑4× efficiency boost after restructuring a legacy system project, reducing team size from 50‑60 people to 4‑5 AI‑coding specialists.

Kuan explains that true waste lies in repeated low‑value interactions; his team built a Spec‑First‑driven AI Coding Harness that embeds expert knowledge into standardized workflow nodes, reducing bugs and rework.

Gu cites academic research that excessive context expansion degrades model performance, reinforcing the need for concise, consistent prompts.

Topic 6: First Move in Enterprise AI – Protocol, Tool, or Cost?

Kuan recommends starting with high‑frequency, high‑pain tools to demonstrate quick value, then gradually introducing protocol and cost governance as usage scales.

Mao adds that company type matters: tech startups can prioritize technology first, while traditional enterprises must calculate ROI before budgeting.

Gu observes that even aggressive AI startups eventually converge on protocol and cost optimization after initial model experiments.

Topic 7: Multi‑Person Requirement Writing, Few Engineers Coding – Can This Model Scale?

Mao describes a pilot where product, development, and testing teams co‑author precise requirements, which are then handed to a small group of AI‑coding experts, dramatically reducing conflicts.

Kuan warns that different roles often have divergent definitions of “high‑quality input,” leading to hidden rework costs.

Gu notes that AI levels the coding skill gap, shifting emphasis toward research and problem‑solving abilities.

Topic 8: Role Re‑shaping in the AI Coding Era

Mao argues that communication may become more critical than pure coding skill, suggesting humanities‑oriented professionals could excel at prompt engineering.

Kuan predicts new hybrid roles (e.g., product‑design‑frontend, dev‑test, or “PDE” – on‑site delivery engineer) to bridge AI‑enhanced workflows.

Gu concludes that team composition will align with business domains rather than static technical labels, emphasizing adaptability and self‑drive.

Topic 9: The Next Hard Problem – Agent Society, Context Base, and Enterprise Architecture

Gu introduces the concept of an “Agent Society,” where agents need identity, reputation, and cross‑organization collaboration, analogous to human societal infrastructure.

Kuan stresses building a sustainable context base with high‑quality knowledge governance, indexing, and on‑demand loading, while also cultivating an organizational culture that embraces AI usage.

Mao highlights the architectural challenge of scaling AI across hundreds of systems, proposing either a universal MCP interface or decentralized multi‑agent coordination, both requiring an AI‑ready blueprint.

Conclusion: One Thing to Focus On

Mao advises picking the single most business‑valuable task for an agent to ensure steady progress.

Kuan suggests first identifying repeatable high‑frequency tasks where AI can add immediate value, then iteratively expanding.

Gu encourages staying engaged with cutting‑edge agent and large‑model research, adopting rather than merely observing the technology.