Agently-Daily-News-Collector — AI Agent 工作流中文教程

Agently Daily News Collector v4

Agently Daily News Collector has been rewritten on top of Agently v4 and now uses:

• TriggerFlow for the end-to-end pipeline
• Agently v4 built-in Search and Browse tools
• structured output contracts instead of the old v3 workflow API

Version constraint: this project requires Agently v4.0.8.3 or newer. The current implementation uses TriggerFlow sub flow to organize per-column pipelines, so earlier v4 releases are not compatible with the workflow structure used here.

The previous Agently v3 project has been archived under ./v3.

Features

• Input a topic and generate a multi-column news briefing automatically
• Search, shortlist, browse, summarize, and assemble stories in one flow
• Save the final report as Markdown under ./outputs
• Keep prompt templates in ./prompts for easy editing
• Keep an independent ./tools layer so search/browse can be replaced without touching the main workflow
• Keep flow construction in ./workflow so orchestration can evolve independently from collector logic

Agently v4 features used here

• TriggerFlow orchestration
• Replaces the old v3 workflow style with a more explicit flow graph (to, for_each, sub flow, branching-ready composition).
• Unlike the old v3 Workflow chain, TriggerFlow here runs columns concurrently and also summarizes picked stories concurrently within each column.
• Meaning for this project: the end-to-end news pipeline is easier to inspect, evolve, and split into chunks without mixing orchestration with business logic, while the parent report flow and the per-column pipeline can now be modeled directly as parent/child flows instead of one oversized chunk.
• Sub flow composition
• The project can now extract a naturally repeated business pipeline, “build one column”, into its own TriggerFlow and invoke it repeatedly from the parent flow inside for_each(column).
• Meaning for this project:
• the parent flow stays focused on report-level orchestration
• the column pipeline can be tested, visualized, and exported independently
• future variants such as “briefing column”, “deep-dive column”, or “regional column” can reuse or derive from the child flow instead of cloning parent-flow nodes
• capture / write_back makes the boundary between parent and child explicit for input, state, and resources
• Structured output contracts
• YAML prompts now define output schema directly for outline generation, news picking, summarizing, and column writing.
• Meaning for this project: much less handwritten parsing glue, clearer interfaces between steps, and easier prompt iteration.
• Built-in Search / Browse tools
• The project now defaults to Agently v4 built-in tool implementations instead of the old project-local helpers.
• Meaning for this project: less custom infrastructure code, and users can still swap implementations through ./tools without rewriting the workflow.
• Runtime resources and state namespaces
• TriggerFlow runtime resources are used to inject logger/search/browse dependencies, while runtime state stores execution data such as request, outline, and intermediate results.
• Meaning for this project: dependency wiring and execution state are separated cleanly, which keeps chunk code thinner and more maintainable.
• Environment-aware settings
• Agently v4 set_settings(..., auto_load_env=True) works directly with ${ENV.xxx} placeholders.
• Meaning for this project: model endpoint, model name, and API key can be switched by environment instead of editing code or committing secrets.