Skip to main content
An objective is one task assigned to one published agent. Cadenya chooses a variation when the objective is created, snapshots the effective configuration, and records everything the agent does as an event. This guide starts in the dashboard so you can see each input and result. The last section shows the same dispatch in TypeScript, Go, Ruby, and cURL.

What you need

  • A published agent with at least one variation.
  • Either a First User Message for this run or a First user message template on every variation the objective can select.
  • Values that satisfy the agent’s system prompt data schema, when it defines one.
The Quickstart builds a suitable Faker-backed agent from an empty workspace.

Create the objective

Open Objectives from the sidebar and select New Objective. Complete the first card:
New Objective form with Conference Attendee Generator selected, Agent default variation selection, an attendee request, and empty system prompt data

The New Objective form with an agent, variation, message, and prompt data

Two fields are conditional:
  • Episodic Key appears when the selected agent has episodic memory enabled. Objectives with the same key share the agent’s corresponding episodic memory layer.
  • Variation is optional. Selecting one pins that variation. Leaving it empty applies the agent’s Random or Feedback Driven selection mode at creation.
Expand Add labels or external ID when the run should carry an identifier from your system or routing labels for webhooks. Select Create Objective. Cadenya opens the new objective immediately while execution continues in the background.

Read the timeline

The Timeline tab is the durable history of the run. It starts with User Message and adds assistant messages, reasoning, tool activity, approvals, compaction, delegation, and terminal events as they occur.
Objective Timeline tab showing attendee generation messages, ten GenerateFake calls, and their results

An objective timeline with messages and tool activity

Use the view selector to switch between:
  • Timeline, which presents the run as a conversation and expandable event cards.
  • Waterfall, which plots event start times and durations on one axis.
The tabs beside it expose the same run from different angles: The Details card records the snapshotted agent, variation, model, temperature, tool counts, and constraints. Editing the agent later does not change this objective.

Understand the state

The status in Details moves through these states: An agent without structured output normally ends a turn in Waiting. That is a live conversation, not a terminal state.

Continue a waiting objective

When the status is Waiting, enter another message in Send follow-up instructions… below the timeline and submit it. The objective returns to Pending, then Running, with its existing event history and configuration snapshot intact. The follow-up contains only the new message. Secrets and prompt data are fixed at objective creation, so start a new objective when those inputs must change.

Submit feedback

Open Feedback and use Submit Feedback after you can judge the result. The dashboard maps its five stars to API scores: Add a comment and select Submit Feedback. If the agent uses Feedback Driven selection, the score updates the posterior of the variation snapshotted on this objective. See feedback and variation sampling for the selection math.
Objective Feedback tab with an Excellent rating and a comment about accurate attendee data

Choose a rating and explain what made the result useful

After submission, the rating appears under Previous Feedback.

Dispatch from code

Set the agent identifier to either its Cadenya ID or an external_id: reference:
Choose a language. Every example creates the same objective and prints its Cadenya ID.
The response starts in STATE_PENDING. Open its ID under Objectives, or stream its typed events from your application.

Objectives from the SDK

Stream, continue, cancel, and score the objective in code.

Get structured output

Define a JSON Schema and finish objectives with machine-readable output.

Approve a tool call

Review a gated tool call in the dashboard or through your own webhook handler.

How objectives work

Understand snapshots, memory, tools, events, context windows, and terminal states.