Scenario-to-Capability Mapping

Languages: English · 中文

When a team asks "can Agently help us build this", the fastest path is to start from the deliverable. What does the business need to receive: fields, a report, an action plan, a controlled system call, an approval-ready artifact, or a recoverable workflow?

Start with the Business Deliverable

Business goal	What the business wants to see	What developers need to add	Agently entrypoint
Turn input into consumable fields	Tickets, forms, risk items, summaries, evidence	Fields, types, required checks, validation, retry, final data readers	Output Control, Model Response
Show progress before the full result	Cards filling in, report sections appearing, staged UI	instant field stream, temporary UI state, final get_data() persistence	Model Response, TriggerFlow Model Integration
Let the model call business capabilities	Query status, read DB, send notifications, control objects	Action schema, visible capability scope, call logs, error semantics	Actions, Action Runtime
Connect MCP, scripts, or specialist tools	Reuse ecosystem capabilities without losing control	Host permissions, redaction, audit; managed execution lifecycle	MCP, Execution Environment
Manage branches, waits, approval, recovery	Know where each step is, why it stopped, who can resume	Execution id, runtime stream, pause/resume, close snapshot	TriggerFlow
Let specialist roles collaborate	Risk, facts, editing, execution roles produce reviewable results	Role agents, Sub-Flow, Workspace topology, structured aggregation	Multi-Role Collaboration
Proactively move tasks or send reminders	Scheduled reports, approval reminders, event-triggered work, IM notifications	Scheduler / webhook / queue boundary with Agently execution	Interaction Layer and Active Tasks
Retain cross-turn evidence	Prior conclusions, reports, downloads, rules, checkpoints	Workspace records/artifacts/checkpoints and goal-budgeted ContextPack recall	Workspace, Context Engineering

Common Scenario Patterns

Customer Support, Operations, and Internal Tickets

Start with output schema. The agent should return fields such as intent, priority, summary, required_action, and customer_reply. A downstream system can route or update tickets from those fields.

Upgrade path:

• Add instant stream when the UI needs a card to fill in while the response is streaming.
• Add Actions when the agent needs to query ticket status, customer profile, billing state, or ownership.
• Add TriggerFlow when the task needs approval, escalation, retries, or human handoff.

Industry News, Research Reports, and Document Collection

The deliverable is usually a report that can keep being edited. A good Agently structure is: topic planning, search/browse actions, candidate selection, summarization, structured section output, and report rendering.

Use TriggerFlow when sources can be processed in parallel. Store source links, summaries, and final report artifacts in Workspace when the team needs traceability.

Document Review, Contract Analysis, and Policy Checking

Do not ask one prompt to "review everything". Split the problem into routing, evidence extraction, issue classification, suggested changes, and approval decisions.

Useful capabilities:

• Output Control for issue records and required evidence.
• Workspace for source snippets, decisions, and report artifacts.
• TriggerFlow for staged review, parallel sections, and human confirmation.
• Model judge or deterministic business checks for acceptance.

Natural-Language Control of Business Objects

The model should not directly execute natural-language instructions. First ask it to produce a structured plan or command object, then let deterministic code check permissions and call Actions.

Typical result fields:

{
  "intent": "update_device_mode",
  "target": "controller-17",
  "args": {"mode": "eco"},
  "requires_approval": false,
  "user_message": "I will switch controller-17 to eco mode."
}

The user-facing message and the internal command are separate fields. That separation makes audit, approval, and rollback possible.

Specialist Workbenches and Complex Toolchains

Professional tools often need structured drafts, deterministic validation, and controlled tool execution. Agently should help the developer separate:

• model-owned reasoning and draft generation,
• deterministic validation and repair,
• controlled Actions or ExecutionEnvironment calls,
• visible runtime evidence for users and operators.

Scheduled, Reminder, and Proactive Workflows

When the system triggers work without a user waiting on the HTTP request, the trigger belongs to scheduler / webhook / queue / worker code. Agently execution should own the reasoning, actions, workflow, and evidence after the task starts.

Capability Packages and Team Practices

When a team has a reusable way to handle a class of work, encode it as guidance, Actions, Skills, or TriggerFlow templates depending on what it actually contains. A Skill should not hide an uncontrolled execution path; it should map back to controlled runtime capabilities.

A Simple Decision Order

1. What fields or artifacts should the business system receive?
2. Does the model need external capabilities?
3. Do those capabilities need permissions, redaction, approval, or sandboxing?
4. Is there a workflow lifecycle: branch, parallel work, wait, resume, or stream?
5. Does evidence need to survive across turns?
6. Does the product need stable process events?
7. What proves the result was correct enough to ship?

Project Evidence

Project type	Scenario evidence	Agently responsibility
Daily News Collector	Topic to Markdown/PDF report	Workflow orchestration, search/browse, structured summaries, report rendering
Talk to Control	Natural-language control of a business object	Action plan, controller boundary, refusal of unsafe operations, process feedback
AgentlyTextParser	Long-document extraction and evidence location	Chunking, field extraction, citation location, aggregation
EDA Agent Final	Specialist toolchain assistant	Structured draft, deterministic validation, controlled tool execution, runtime evidence