A grove of tall trees rising from a green valley at dusk

Agents

Give agents
real work.

Connect tools, delegate to specialists, and carry memory between tasks. Try new approaches with variations, then let real feedback guide what runs next.

A variation can change the model, prompt, tools, memory, and specialist agents. Explore a different approach behind the same agent your product calls.

Default

Anthropic Claude Sonnet [Latest]

Canary

OpenAI: GPT-5.5

new variation

Canary

Model

GPT - 5.2

Created

2 days ago

Assignments

  • reroute shipment tool
  • update_eta tool
  • Dispatch API tool set
  • Customs Broker sub-agent

Memory layers

Carrier Playbook - SLA Policies

System prompt

You are Meridian’s shipment-exceptions agent. When a delivery stalls, reroute it via the dispatch API...

Choose a different model or provider for each variation. Compare approaches while your product keeps calling the same agent.

Default

Anthropic Claude Sonnet [Latest]

Canary

OpenAI: GPT-5.5

new variation

Assign tools, tool sets, sub-agents, and pools per variation. Give each approach its own capabilities and discover tools on demand.

Enable progressive tool discovery

Max tools per search

1-10

5

Search hints

up to 5

delays, reroutes, customs

Rerank threshold

0.0-10 [Leave blank to skip reranking]

Tool schemas stay out of the context window until the agent asks for them — only names ride along, so every request gets smaller.

Duplicate a variation and try a new approach. Score the outcomes to help Cadenya choose which variation handles the next objective.

Feedback

Search comments All sentiment All variations
Created Comment Variation objective
+1.00 Rerouted before the SLA breach-recipient never noticed. Canary | GPT-5.2 view objective
+0.50 Caught the customs hold and updated the ETA proactively. Default | Claude Sonnet 5 view objective
+0.50 Good reroute, but notified the recipient twice. Canary | GPT-5.2 view objective
-0.25 Held at facility when a reroute was available Canary | GPT-5.2 view objective
+0.75 Escalated the frozen-goods lane correctly Default · Claude Sonnet 5 view objective

There’s a
webhook
for that

Receive webhooks for everything
your agent is doing. Everything.

A flowering bonsai on a floating island, its roots trailing below 
POST /webhooks/cadenya HTTP/1.1
Host: your-app.com
Content-Type: application/json

{
  "type": "objective_event.tool_result",
  "data": { ... }
}

Run your agent

// Template: "Track order {{ orderId }}"
await cadenya.objectives.create({
  agentId,
  firstUserMessageData: { orderId: "1234" }
});

Tag it to your customer

await cadenya.objectives.create({
  agentId,
  tenant: { id: "acme" },
  subject: { id: "user_123" },
  firstUserMessage: "Where is my order?"
});

Track input, output, cached, and reasoning tokens over time. Filter by agent, compare models, and see token spend alongside cache hits.

Token consumption

  • Input
  • Cached input
  • Output
  • Reasoning
60K45K30K15K0 MonTueWedThuFriSatSun

Define a JSON Schema for your agent’s output. Turn finished work into structured data your application can use.

Structured output

{
  "orderId": "1234",
  "status": "delayed",
  "action": "notify_customer"
}

JSON Schema → Agent result

Follow objectives, loops, tool calls, and feedback by agent and variation. Spot failure hotspots and see where work needs attention.

Agent health

VariationObjectivesLoopsTokens
Default1696284K
Canary1272192K

Map the models you use to their replacements, even across providers. Move matching agent variations together, without editing each one.

Model swap

From · OpenAI

GPT-4.1

To · Anthropic

Claude Sonnet

Perform swap

Give your agent
a team.

01 / Tools

Put tools to work.

Give a variation individual tools or entire tool sets. Connect the actions and systems it needs to finish the job.

02 / Sub-agents

Call a specialist.

Let an agent delegate a task to another agent. Build specialists once and put their expertise to work across your team.

03 / Agent pools

Route the request.

Group specialists into a pool. Routing instructions help choose the agent best suited to each request.

A few things
to know.

What can I change between variations?

The model, prompt, tools, tool sets, sub-agents, pools, memory layers, and runtime settings can all differ.

How do pools choose an agent?

A pool routes a request to one of its member agents using your routing instructions and the context supplied with the request.

Can agents remember across tasks?

Yes. Enable episodic memory and supply the same episodic key across objectives. All variations share that key’s memory layer.

A sprawling bonsai on a floating island trailing roots

Put your agents
to work.

Start with one. Give it tools, a team, and room to grow.

Create an agent See API docs