Agents

An agent is the AI layer of a scene. It's what reads the prompt, picks tools, calls them, decides what to do next, and produces output.

There's no special "agent" object you configure in most cases — the scene comes with one by default. The interesting choices are which model the agent uses and which tools it has access to.

The model is your choice

Anthropic, OpenAI, OpenRouter, Gemini, Amazon Bedrock, SAP AI Core. By default you run on managed usage with a hard spend cap; on Enterprise you connect your own keys once at the workspace level, and every scene uses your providers directly.

You can switch models per conversation. Use Claude for the planning, GPT for a specific call, Llama via Bedrock for cheap bulk work. The agent loop doesn't care which model you pick.

If you don't have keys yet, we include some Claude usage on the free tier to get started. You'll outgrow it quickly if you're using Daslab seriously, but it's enough to evaluate.

More on BYOK.

Tools come from integrations

A tool is a function the agent can call. Tools live behind integrations — when you connect Gmail, you get tools like gmail.search_threads, gmail.send, etc. When you connect Postgres, you get postgres.query, postgres.execute_sql. When you connect a deeper system like S/4HANA, you get tools for OData queries, BAPIs, and live ABAP execution.

The agent picks tools by name and parameter shape. It can chain them — call one, use the output to call another, join the results, draft a write. You see the chain in the job trace.

Tools that write (mutating tools) trigger an approval gate before they execute. Read tools run freely.

When no pre-built tool fits, the agent can write its own code — JavaScript, SQL, shell, and bespoke languages where needed — and run it in an isolated sandbox. Output is captured, the sandbox is torn down.

Anatomy of a run

A single agent run looks like:

  1. Read the user prompt + scene context.
  2. Decide what to call.
  3. Call it. Read the output.
  4. Decide whether to call more, ask the user, or write to a system.
  5. If writing: stop and ask for approval. Show diff.
  6. Repeat until done.
  7. Produce final output. Update cells if applicable.

The whole loop is recorded — every tool call, every model decision, every approval gate — and every conversation has a permalink to that record.

One agent per scene

One scene, one agent. It carries the scene's toolset and model choice; separate workflows get separate scenes.

What's next

Updated 2026-04-30