Building with Agents

Using AI to author graphs without letting it near the hardware.

Who this is for: Engineers adopting AI assistance who need results they can defend.

The graph as a specification

What a typed dataflow graph gives an agent to work from, and you to check.

1 min

Older IDE Filmed on the IDE as it was in late August 2026. The editor has changed since, so some screens look different. The written walkthrough is current.

Describing intent as ports and types.
Letting an agent propose a graph.
Reading the proposal critically.

Outcome: You can brief an agent in a form that can be checked.

Verification is the job now

What the type system, the graph rules and tests prove — and what they do not.

1 min

Older IDE Filmed on the IDE as it was in late August 2026. The editor has changed since, so some screens look different. The written walkthrough is current.

Structural validation before anything runs.
Compiling the emitted Rust and testing it.
Knowing which guarantees you actually hold.

Outcome: You can state precisely what has been proven about generated work.

Keeping the model outside the loop

The architecture that lets you use an agent and still sign off on the result.

1 min

Older IDE Filmed on the IDE as it was in late August 2026. The editor has changed since, so some screens look different. The written walkthrough is current.

Propose, validate, execute — and who does each.
Why a non-deterministic model cannot sit inside the safety function.
Designing a workflow you could show an auditor.

Outcome: You can use agents on work that has to be defensible.

Open the IDE