Building Agentic Workers in the Designer

The Greentic Designer lets you build an agentic worker visually, without hand-writing an AgentConfig or packaging tools yourself. It produces the same artifact the runtime runs: a .gtpack containing the worker’s config.

There are two surfaces, and they share one backend:

• DW Composer Studio (/dw-composer) — a form: fill in the worker section by section.
• Chat ✨ Auto-fill — describe a goal in plain language and let a composer propose the whole worker, which you then refine in the Studio.

DW Composer Studio (the canonical path)

Open the dashboard → Agentic Worker tab → New Project. The Studio is a stacked form:

Section	What you set
Basic	Display name, description, tenant, locale, vertical
Instructions	The goal, the system prompt, the opening message
LLM	Provider + model (e.g. OpenAI gpt-4o-mini)
Tools	The tool extensions the worker may call
Memory	Short-term / long-term memory providers
Channels	Messaging channels (webchat, Slack, …)

A sticky Manifest Preview shows the worker taking shape. When you’re happy, click Build — the Studio runs a pure-function mapper (DwFormState → AnswerDocument) and packages a .gtpack. Download it, or Publish it to the Greentic Store / Deploy it to a runtime.

注意

The Studio form path is self-contained — it does not need a composer extension. The form maps directly to the worker’s AgentConfig, the same shape described in Agentic Workers.

Effective prompt — seeing what the model actually gets

Beside the Manifest Preview sits the Effective prompt panel, collapsed behind a Show what the AI receives toggle. It is read-only, and it shows the system message a deployed worker would receive — not an approximation — plus one row per bound tool with its description and usage note.

Why it exists depends on which instruction mode the worker uses:

Mode	What the model receives	What the panel adds
Custom	Your Instructions, verbatim	Little — the prompt is what you typed
Managed	A platform-authored backbone with your instructions slotted in	The point of the panel

In Managed mode the platform supplies a standard structure — an identity line, a workflow-and-escalation paragraph, and the guardrail framing — and you only write the business-specific layer. Because the result is one continuous prompt, it is otherwise impossible to tell your sentences from the platform’s. The panel therefore renders it as labelled blocks:

• identity — the “You are …” line, from your display name and tone
• added by the platform — the fixed backbone you cannot edit
• your instructions — the layer you wrote

If you have not written instructions yet, that block says so rather than sitting blank. A Copy button puts the whole system message on your clipboard.

警告

The preview is composed server-side, through the exact same code path that builds a deployed worker’s config — so it cannot quietly drift from reality. Note what it does not show: knowledge bases are not prompt text. On the agent path they are retrieved per turn at runtime, so an attached knowledge base correctly appears nowhere in this panel.

Chat ✨ Auto-fill (compose from a goal)

In the Studio, type a goal (e.g. “A read-only network operations assistant for our telco”) and click ✨ Auto-fill from goal. The Designer:

1. Asks the LLM for 0–3 clarifying questions (/api/dw/composer/plan-questions).
2. Sends your goal + answers to a composer — a vertical-specific extension that proposes the worker (compose_digital_worker).
3. Projects the result into the Studio form and shows a per-field diff so you can review before applying.

Composers are design extensions advertising the greentic:dw-composer/composer capability. The Designer discovers every installed composer at startup and selects one by vertical at compose time. The bundled set covers default, telco, banking-id, cs-bot, healthcare, public-sector, retail, and template-builder.

Prerequisites for Auto-fill

1. Install composer extensions (v1.3.0+ — older apiVersion v1 builds will not load):

   gtdx install greentic.dw.composer.default
   gtdx install greentic.dw.composer.telco

   They land in ~/.greentic/extensions/design/. On its next start the Designer logs how many it discovered, e.g. Discovered 8 DW composer(s) from 8 advertised extension(s).

2. Configure an LLM. Standalone: --openai-api-key / GREENTIC_LLM_API_KEY (or pass --llm-provider). With an admin backend, the agentic_worker_composer role provides it.

警告

If no composer is installed, Auto-fill returns a structured “no composer” result — use the Studio form directly. (Discovery is LLM-free; only the actual composition calls the LLM.)

From worker to running flow

A Designer-built worker is the same artifact as a hand-authored one, so once it’s packaged it runs exactly as described in Agentic Workers: a dw.agent node in a flow hands user messages to the worker’s Plan→Act→Observe loop.

Step	Where
Build .gtpack	Studio Build button
Publish to Store	Publish (POST /api/dw/publish)
Deploy to a runtime	Deploy (POST /api/dw/deploy, needs an admin endpoint)
Test before shipping	the Studio test-chat playground