Deploying AI Agents: Protocols, Costs, and Evolution from Demo to Production

On June 25, 2026, the DataFun community hosted a deep‑dive dialogue on enterprise‑grade AI agents. Moderator Gu Yongfeng (FoundationAgents) and two frontline technical leaders—Kuan Yuping (Shenzhen Value Network) and Mao Zhuo (Country Garden Services)— explored three core questions: protocol collaboration, self‑evolution ability, and cost control.

Topic 1 – Overlooked Shortcomings – Mao emphasized that an agent must deliver clear business value; otherwise, massive token consumption becomes unsustainable. He estimated that scaling an agent could cost tens of millions of tokens annually, prompting a ROI calculation. Kuan highlighted engineering gaps: unstable context between product, development, and testing, lack of traceability, and difficulty governing outcomes, concluding that enterprises need a full‑stack, process‑oriented agent framework rather than a smarter chatbot.

Topic 2 – Protocol vs. Spec‑First – Gu introduced the Foundation Protocol (FP) as a solution for inter‑agent communication, using the analogy of a manager juggling many contacts. Kuan positioned the protocol as the “bottom‑level infrastructure” while the Spec‑First process serves as the immediate entry point for workflow integration, describing the protocol as a long‑term foundation and the process as the practical landing pad.

Topic 3 – Tool vs. Protocol Acceptance – Gu argued that tools are the entry key, likening them to HTTP/TCP‑IP where developers use packaged SDKs without worrying about handshakes. Kuan added that early‑stage tools provide quick wins, but without a unified protocol, rapid tool proliferation becomes a maintenance nightmare. Mao stressed that business stakeholders need simple, reliable outcomes; otherwise projects are abandoned.

Topic 4 – Token Cost Explosion – Mao noted that token waste stems from both technical and managerial sources. Different IDE/CLI tools vary in token efficiency, and individual communication styles amplify consumption. Kuan identified three waste sources: repetitive injection of large context blocks, unlayered knowledge, and lack of reusable process assets, describing large models as “a fish with a seven‑second memory and a Swiss‑army‑knife of capabilities.”

Topic 5 – Cost Reduction & Efficiency Gains – Mao shared a case where monthly costs fell 88 % and development speed increased 3‑4× by restructuring a legacy system project: a small AI‑coding team handled code generation while product, development, and testing focused on precise requirement writing and review. Kuan’s Spec‑First‑based AI‑Coding Harness reduced redundant interactions, lowered bugs, and improved knowledge reuse.

Topic 6 – Choosing the First Move – Kuan recommends starting with tools that address high‑frequency pain points, then gradually introducing process and protocol layers as usage scales. Mao adds that the choice depends on company type: startups may prioritize technology first, whereas traditional enterprises must validate business value and ROI before heavy investment.

Topic 7 – Role Re‑shaping in the AI‑Coding Era – Mao described a pilot where cross‑functional squads (product, dev, test) co‑author precise requirements, then a few AI‑coding experts feed those to the model, dramatically reducing conflict. Kuan highlighted divergent understandings of “high‑quality input” across roles, stressing the need for clear standards and pre‑defined constraints. Gu observed that AI flattens coding skill differences, shifting value toward research, problem‑solving, and communication abilities.

Topic 8 – Future Hard‑Knocks – When asked about the next bottleneck, Gu introduced the concept of an “Agent Society” requiring identity, reputation, and cross‑organization collaboration infrastructure. Kuan pointed to sustainable context bases (knowledge, memory, graph, vector logs) and cultural adoption as critical. Mao emphasized enterprise‑wide AI‑ready architecture, describing two paths: a super‑agent interfacing all systems via MCP, or decentralized agents coordinated through multi‑agent scheduling.

In conclusion, the panel agreed that moving an agent from “demo‑ready” to “production‑ready” is not a pure model‑performance issue; success hinges on a triad of engineering systems, management processes, and underlying protocols that together enable scalable, cost‑effective deployment.