How Daslab works

Daslab is a workspace where AI agents work in the systems you already run — your inbox, your ERP, your servers, your repos. You give direction in plain language. The work happens in a place you can look at.

That place is a scene. Every workflow becomes a scene — its data, its tools, the agent that runs it, and the history of everything it did. A workspace holds many scenes: one chasing overdue invoices, one watching a server, one preparing the Monday report. The rest of this page walks through one of them.

A scene is where one workflow lives

Everything the workflow needs lives inside. Take the invoice scene: it holds the inbox the invoices arrive in, the list of who hasn't paid, the reminders waiting to go out, and the agent that keeps it all moving. Open it and you see the state of the work — not a chat log, the work itself.

Every cell looks at an asset

The things in a scene — the inbox, the repo, the draft that hasn't been sent — are assets: governed, gated, audited. Cells show them: point a cell at an asset, pick a view and a size, and it stays connected — the inbox cell doesn't hold a copy of your inbox, it looks at the inbox.

Cells work like a spreadsheet, projecting richer things. In the invoice scene, one cell shows the invoices arriving in the inbox. A second derives the ones unpaid past thirty days. A third shows that list to your team. When a payment arrives, every cell that looks at it follows. Nobody refreshes anything.

An agent runs the workflow, on the model you choose

The agent works inside the scene. You tell it what the workflow should do, and it reads the cells, acts on the assets behind them, and builds new ones as the work requires. Ask it on Thursday why a reminder went out and it answers from the cells in front of it.

Agents run on the model you choose: Anthropic, OpenAI, Amazon Bedrock, SAP AI Core, Gemini, OpenRouter — managed usage by default, your own keys on Enterprise. And you can reach your scenes from tools you already work in: Claude Code, Codex, Cursor, or any MCP client.

Real-world actions wait for approval

When the work means touching the real world, the action waits at a gate. The reminder exists in the scene as a draft first; you read it before it goes. Actions you trust go on auto; the default is that you look first.

Every run is recorded end to end

What the agent did, which cells changed, what it cost — open any run and read it end to end. When something looks wrong on Thursday, you don't reconstruct Tuesday from memory. You read it.

100+ native integrations plus any MCP server

Gmail, Slack, Microsoft Teams, and SAP S/4HANA and SuccessFactors are among the native ones; anything that speaks MCP plugs in alongside them.

Hard cap, prepaid, $0.25–$0.75 a typical run

You set a hard cap and prepay the allowance. There are no overage invoices, ever.

The concepts in detail

  • Scenes — the container: what a scene holds and how teams share it.
  • Assets — the things you govern: identity, gates, audit.
  • Cells — how scenes show assets: projections, sizes, views.
  • Agents — the loop inside a job.
  • Jobs — one conversation, its tool calls, its trace.
  • Programs — the agent's plan as an artifact: save it, run it again.
  • Approvals — how writes are gated.
  • Sandbox — when no tool fits, the agent writes code and runs it isolated.
  • Providers — the integration layer asset types come from.
  • Blueprints — pre-built scenes to deploy and adapt.
  • Provider scenes — providers authored as data.
  • Commit chain — the hash-linked history under every scene.
  • Traces — the OpenTelemetry record every run leaves.
  • Audit — the query layer over the chain.

Start with a workflow you already do by hand

Set your first scene: the invoices, the report, the check you run every morning — pick one and put its pieces in cells. Get started walks the whole loop end to end in five minutes, against your own tools.

Once the concepts are clear, the operating docs cover running Daslab in your environment: auth, BYOK, data residency. For the design thinking behind these primitives — Apple WidgetKit, Tufte, Christopher Alexander, Alan Kay, Bret Victor — see influences.

Updated 2026-07-18