You stay in control. By design.
Autonomy without governance is just a black box that emails people. Everything below is a screenshot or recording of the real product — the same views you get on every run.
Approval gates — you approve before anything goes out
Any step that sends something outward gets an approval step in front of it — enforced by the platform, not by convention: a worker without the gate can't be saved, and the engine refuses to run an ungated send. That covers workers composed for you, hand-written ones, installed ones, and scheduled runs.
- The run parks at the gate and waits for your explicit approve — no approve, no send.
- The gate is written into the worker itself — a numbered step in the run, right before the send.
- Approving is a one-click, one-time action — a duplicate or concurrent approve is rejected, not double-sent.
Every run is a step list you can read
No log-diving. A run is a numbered, plain-English list of steps — what each one did, how long it took, what went in, and what came out.
- Status, duration, inputs and outputs on every step — expandable in place.
- Runs that need your approval float to the top of your roster and deep-link straight to the run that's waiting.
- Sub-workers are walkable the same way: a gate deep inside a composition still bubbles up to the top-level run.
Every AI call is on the record
Each step keeps a trace of every single AI call it made: which model, how many tokens in and out, retries, latency — and the exact prompt and response text, expandable right in the step detail.
- Per-call rows, not aggregates — you can audit the one call that mattered.
- The same trace covers your own keys and platform keys alike.
- Recorded below: a real production run's trace, opened live.
The quieter guarantees
Three more things the platform does for you on every run.
Your keys, your call
Bring your own AI keys (Gemini or Anthropic) and your AI calls run at zero platform credit cost — the keys are stored in the platform vault and used only for your runs. No keys? Platform keys work out of the box, metered transparently against your credit balance.
Grounded in live sources
Research agents search the live web through a real Google-results search primitive — prospects, buyers, and signals come from pages that exist today, not from model memory. Search results carry their sources so you can check them.
Out of credits? It stops loud.
If your balance runs out, the run stops with an explicit credits-exhausted error before the AI call is made, and tells you exactly where it stopped. Top up or connect your own keys, then run again.