Steps

Steps are the work a workflow does. They run in the order you write them. This page shows each step type with a small example.

Step type	What it does
read_data	Reads work items from Jira or Azure DevOps.
ai_task	Runs an AI agent with a prompt. No repository is checked out.
coding_task	Runs an AI agent inside a cloned repository. The agent commits and opens a pull request.
send_teams_card	Posts a card to a Microsoft Teams channel.
send_message	Sends a message to a Microsoft Teams chat.
transform	Picks one value out of a JSON output.
echo	Returns its own settings. Useful for testing.

Every step needs a name and a type. Most steps also have a config block for their settings.

read_data

Reads items and returns them as a list. It needs a connection.

- name: fetch-stories
  type: read_data
  provider: jira
  connection: Jira
  config:
    query: project = DEMO AND status = "To Do" ORDER BY priority DESC

For Jira, query is JQL. For Azure DevOps, query is WIQL, and organization and project are also required:

- name: fetch-work-items
  type: read_data
  provider: azure_devops
  connection: Azure DevOps
  config:
    organization: your-org
    project: your-project
    query: SELECT [System.Id] FROM WorkItems WHERE [System.State] = 'New'

When more steps follow, each item in the list gets its own child run. See Fan-out.

ai_task

Runs an AI agent. config.prompt is required.

- name: summarize
  type: ai_task
  config:
    prompt: |
      Summarize this ticket in three bullet points:
      {{steps.fetch-stories.output}}

Two things to know:

• Tools are off by default. Without a system_prompt, the agent is told to answer with text only and not to run commands. If the step must call an API or write files, set system_prompt under config.llm_config.
• Connections become tokens. List connection names under connections. The agent receives one access token for each. The token name is the connection name in capitals, with _ACCESS_TOKEN added. For example, Jira becomes JIRA_ACCESS_TOKEN.

- name: comment-on-ticket
  type: ai_task
  connections: [Jira]
  config:
    llm_config:
      system_prompt: |
        Use your tools. Call the Jira REST API with curl and post the comment yourself.
    prompt: |
      Post a one-line comment on this ticket: {{steps.fetch-stories.output}}
      Use the token in JIRA_ACCESS_TOKEN. Never print the token.

An ai_task cannot have a repos setting. Use coding_task for work in a repository.

To ask a person a question from this step, see Teams cards.

coding_task

Runs an AI agent in a cloned repository. It needs config.prompt, config.repos, and one connection to GitHub or Azure DevOps.

- name: fix-typo
  type: coding_task
  connection: GitHub
  config:
    repos:
      - https://github.com/your-org/your-repo
    prompt: |
      Fix the spelling mistakes in README.md.

After the agent finishes the change, it commits, pushes a branch named anthara/<short-name>, and opens a draft pull request.

Commits are made under the workflow’s slug. For example, the slug ci-fix commits as Anthara CI-FIX.

send_teams_card

Posts a card to a Teams channel. All three settings are required.

- name: announce
  type: send_teams_card
  config:
    title: Nightly check finished
    message: The nightly check finished with no findings.
    teams_channel_id: "19:replace-with-your-channel-id@thread.v2"

send_message

Sends a message to a Teams chat through a Microsoft Teams connection. The message is sent as the person who owns the connection.

- name: notify-chat
  type: send_message
  provider: microsoft_teams
  connection: Microsoft Teams
  config:
    chatId: "19:replace-with-your-chat-id@thread.v2"
    message: The workflow finished.

transform

Picks one value out of a JSON output. input is the JSON text. path is the field to pick.

- name: pick-key
  type: transform
  config:
    input: "{{steps.summarize.output}}"
    path: issueKey

Without path, the step returns the whole input.

echo

Returns its own config. Use it to check a template without starting an agent.

- name: show-input
  type: echo
  config:
    message: "Ticket is {{inputs.ticketKey}}"

Pass data between steps

Use {{ }} inside any text in a step’s config.

Template	Gives you
{{inputs.ticketKey}}	The trigger input named ticketKey.
{{steps.fetch-stories.output}}	The full output of the step named fetch-stories.
{{steps.summarize.output.issueKey}}	One field of a JSON output.
{{steps.summarize.output.items[0].id}}	One field of a list item in a JSON output.

Templates are filled in when the step runs. If a template names an input, a step, or a field that does not exist, the step fails.

An ai_task output is stored as JSON. If the agent’s answer is not JSON, Anthara wraps it as {"output": "<the answer>"}. To read fields from a later step, ask the agent to end with a single JSON object.

Model settings

llm_config sets limits for the AI agent. Put it at the top level to cover every agent step. Put it under a step’s config to change it for that step only.

llm_config:
  max_turns: 40
  max_budget_usd: 5

Key	Meaning	If you leave it out
model	The model the agent uses.	claude-sonnet-4-6
max_turns	The most turns the agent may take. A whole number above zero.	300
max_budget_usd	The most the step may spend, in US dollars. A number above zero.	50
effort	How much reasoning effort to use, such as low, medium, or high.	The model’s default
system_prompt	Standing instructions for the agent.	See ai_task above

Any other key is rejected when you save the workflow. This protects you from a misspelled limit that would otherwise be ignored.

If your organization routes workflows through the LLM Proxy, model must name a model that an admin registered. See Models.