How Yidian Tianxia Built an AI Middle Platform for Autonomous Ad Agents

01 Yidian Tianxia's AI Architecture

Yidian Tianxia is a global marketing technology company serving outbound enterprises and brands, connecting over 1.6 million media outlets. Its business spans three lines: Cyberklick for integrated marketing on Google, Meta, TikTok; Yeahmobi/GridAds for digital advertising; and zMaticoo for programmatic advertising.

The architecture embeds AI inside business, not as a side plugin. The top layer runs business-specific agents (Cyberklick, Yeahmobi, GridAds, zMaticoo). The middle layer provides a unified AI middle platform and data middle platform: business data flows into the data middle platform, then the AI middle platform analyzes it to guide the next business actions. The bottom layer is an intelligent multi-cloud management platform.

Data accumulation metrics: 100+ daily active agents, 130+ MCPs, 1,500+ skills, Q2 token usage up 3x quarter-over-quarter. This production-scale footprint is called an "Agent dark factory" running 7×24.

02 Why Integrate Agents into Ad Business?

Agents address four typical ad pain points: long delivery chains, many agent integrations, anomaly detection relying on experience, and creative adaptation across platforms.

Value evaluation focuses on two questions: can the agent follow the process, invoke capabilities, complete the task, and return results? At core configuration nodes, does it know to pause for human judgment?

Development path: toolization → agentization → collaboration → autonomization. Shared capabilities (models, tools, skills, observability) must not be rebuilt per business line; the unified middle platform solves this.

03 GridAds Smart Investment Engine Case

GridAds handles mid-to-long-tail advertising. The six-step delivery chain: competitor research, channel engagement, delivery configuration, monitoring, strategy tuning, settlement reconciliation, and result feedback. "Missing a window burns real money."

The chain is decomposed: which parts go to agents, which need human confirmation. Overall AI-driven, human-assisted. Example: channels aggregated technically; agents match combinations per client needs. Clients enter agent groups for auto daily/weekly reports and issue responses, but external output still requires human review. Model hallucinations cannot directly touch clients.

Evolution: single agent → internal linkage → multi-agent collaboration → real-time linkage with client and ad systems.

Three collaboration levels:

Client-side ops: internal test analysis, todos — closed within role.

Media-side ops: internal config, monitoring, tuning — closed within role.

Cross-business: client-side and media-side agents collaborate to take over full delivery, with human approval when needed. This is the hardest.

In the closed loop, agents run; humans confirm material approval, strategy sign-off, settlement review, core config. "Agents run process, humans guard gates. Gates will become fewer."

Three-layer structure (cognition, construction, operation) maps to eight steps. Key is "Skillization": turn veteran experience into capabilities with defined inputs, steps, standards; feed every false alarm back into rules and evaluation; next round more accurate; experience accumulates continuously.

Results: anomaly diagnosis from hours to minutes. More importantly: every step traceable, new hires invoke skills directly, agents deliver verifiable service to clients. "Revenue not just efficiency, but managed service from unverifiable to verifiable."

04 Building the Unified AI Middle Platform

Not just GridAds — all three business lines need agent+management integration. Avoiding duplicate investment, shared capabilities are consolidated into the Yidian Tianxia AI middle platform with five layers tightly bound to business.

Model layer: Multi-model access and routing via AI Hub. Anomaly diagnosis uses stronger reasoning models; daily summaries use cheaper models. Auto-switch by complexity.

Capability layer: Tool library — skills, MCPs, knowledge bases. Delivery strategy skills, channel interfaces as MCPs; tools that actually execute in business processes.

Agent layer: Intent classification (parameter tuning vs reconciliation) routes to appropriate agent; not all tasks go to one assistant.

Collaboration layer: Multi-agent orchestration with human confirmation at key nodes (strategy sign-off, settlement review).

Observability & Security layer: Uses Alibaba Cloud AgentLoop for trace observability and evaluation; security and policy guarded by AdClaw. Alibaba Cloud capabilities continuously evolve the middle platform.

AgentCore helps build and govern business processes; AgentLoop provides runtime observability for evaluation and optimization, helping agents enter production with visibility, evaluability, improvability.

05 Observability and Tuning Agents

"Visible" to clients means: budget tracking, every step traceable, not just post-hoc reports. AgentLoop opens the black box: captures reasoning, tool calls, memory, sub-agent trajectories; each call's input, output, latency traceable. Divergence or failure can be located on the graph.

Three-step flywheel: leave traces → define good/bad → feed back to production.

Leave traces: Data ingestion (traces, logs, datasets); depth depends on instrumentation completeness. Structured trajectories: tasks, tool calls, agent trajectories together. Standardized execution traces.

Audit & approval: Via AdClaw business process and AgentLoop, combine run trajectories with human roles. Session-level audit event stream. From GridAds: threshold judgment and parameter tuning have trace audit, orchestratable, replayable.

Define good/bad (three steps):

Evaluation assetization: historical trajectories → Golden Set and Bad Cases. AI evaluator + business rules score. Production evaluation tasks and prebuilt evaluators show score distribution bias.

Regression evaluation as gate: skill, prompt, model version changes must pass before release.

Continuous optimization: prompt changes, skill updates, model routing/threshold adjustments require human confirmation before feedback. Agents cannot decide alone.

Four key closed-loop metrics: diagnosis accuracy, steps to locate, false positive rate, token cost efficiency.

06 Multi-Agent Collaboration

Alibaba Cloud AgentCore supports cross-business collaboration: models, skills, MCPs, agent teams, credentials, permissions. Who can call what, with what identity — AgentCore handles the technical plumbing. AgentLoop runs on AgentCore. Runtime visibility, evaluation, diagnosis, human-confirmed skill changes. Stable base enables the flywheel.

Three collaboration cases: ops tickets, GridAds cross-business, zMaticoo multi-business orchestration. Evaluation criteria: orchestration correctness, context continuity, full traceability, human approval for high-risk, token cost-effectiveness.

Results: AI covers ~50% core execution work, ops per-capita efficiency up 40%+. Next targets: task execution closure rate to 80%, ops efficiency coverage >50%.

Conclusion: Key to agent scaling is not just model strength, but: business processes collaborative, runtime observable, capability effects evaluable, high-risk actions governable. Ad lifecycle is long; agent navigation just starting.