Zum Inhalt springen
greentic.ai Greentic Docs

gtc worker

Dieser Inhalt ist noch nicht in deiner Sprache verfügbar.

gtc worker builds an agentic worker — of any type — into a runner-loadable .gtpack, entirely from the command line. It is the CLI counterpart to the Designer composer: you describe the worker in a WorkerSpec file (YAML or JSON), and gtc worker build produces the same pack the Designer would.

Commands

Section titled “Commands”
CommandPurpose
gtc worker init <single_turn|agent_graph|deep_worker>Write a starter WorkerSpec for that type
gtc worker newInteractive wizard → a WorkerSpec → a built .gtpack
gtc worker build <spec.yaml> [-o out.gtpack]Build a .gtpack from a WorkerSpec
gtc worker validate <spec.yaml>Validate a WorkerSpec without building
Terminal window
gtc worker init single_turn -o triage.yaml   # scaffold
# …edit triage.yaml…
gtc worker validate triage.yaml               # check
gtc worker build triage.yaml -o triage.gtpack # build
gtc start ./triage.gtpack                      # run

The WorkerSpec

Section titled “The WorkerSpec”

One file describes the whole worker. The interactive gtc worker new produces the same format, so the wizard and the file path are interchangeable.

apiVersion: greentic.ai/v1
kind: single_turn            # | agent_graph | deep_worker
name: support-triage
llm:
  provider: openai
  model: gpt-4o
  credential_ref: llm-openai
instructions: |              # system prompt
  You are a support triage agent. Resolve the customer's intent, then route or answer.
tools: [ web.search, hubspot.contacts ]
memory:
  short_term: true
  long_term: { provider: chronicle, credential_ref: chronicle-main }
knowledge:                   # RAG — see the caveat below
  provider: chronicle
  embedding: { provider: openai, model: text-embedding-3-small, credential_ref: embed-openai }
  top_k: 5
  documents: [ ./kb/policies.pdf, ./kb/faq.md ]
guardrails: [ pii-redact ]
  • kind selects the worker type. The type-specific block (below) is required only for agent_graph / deep_worker.
  • tools is the allow-list of tool ids the worker may call.
  • memory and knowledge carry the same settings as the Designer — see Agent Memory and Agent Knowledge.
  • documents are read locally, text-extracted, chunked, and baked into the pack’s corpus.

agent_graph

Section titled “agent_graph”
kind: agent_graph
# …llm / instructions / tools as above…
agent_graph:
  coordinator: { instructions: "Route to the right specialist." }
  specialists:
    - { name: billing, instructions: "Handle billing questions.", tools: [stripe.lookup] }
    - { name: tech,    instructions: "Handle technical questions.", tools: [docs.search]  }

deep_worker

Section titled “deep_worker”
kind: deep_worker
# …llm / instructions / tools as above…
deep_worker:
  iterationBudget: 8         # long-horizon plan → reflect → delegate loop
  reflection: true
  delegation: false

What build produces

Section titled “What build produces”

gtc worker build emits a runner-loadable .gtpack: a manifest.cbor, a dw-agents.json carrying each agent’s config (llm, tools, memory, knowledge, guardrails), a flows/main.ygtc with the type’s executing node (dw.agent / dw.agent_graph / operala.call), and — when knowledge.documents is set — the baked knowledge corpus. The pack runs with gtc start.

See also

Section titled “See also”