Human on the gates.
Not in the loop.

Autonomous end-to-end — you open the gates that matter.

Prompt engineering
→ graph engineering.

The work has a shape you can see, edit and version.

The graph grows at runtime — one agent per task.

Every run
has a price tag.

Per step. Per model. Per agent. Before the invoice.

Spot the step that burns 60% of the budget.

Set a budget — the run stops itself.

Cloud. On-prem. Local.
One pipeline.

Judgment is expensive — buy it per token.

Bulk work — run cheaper on-prem.

Chores are free — run them locally.

Runs live
where the team talks.

Start, question, nudge, approve — from chat.

Replies flow back into the run.

Not one laptop. A shared, visible activity.

Policy travels
with the run.
Not with the developer.

Pick a guardrail set when you start — Normal, Strict, or your own.

Every agent the run spawns gets the same rules — denies + scrubbed env.

Lower scopes can add rules, never remove them. The set is on the run's record.

Worst case:
a thrown-away clone.

Every run works in its own cloned tree.

A bad run is discarded with zero blast radius.

Many runs, one repo, no collisions.

Ask Worca.
It was there.

A conversation over every run, log, cost and diff.

Answers grounded in actual run data.

An audit trail you actually use.