PIPELINES & GATES
Nothing ships without a human.
Multi-stage pipelines that stop mid-flight until a second person approves. Not a checkbox before the run starts — a pause at the step that matters, with the exact change in front of the approver.
THE MANUAL WAY
Why approvals usually fail.
THE RUBBER STAMP
Approval happens in a chat message before anyone knows what the change will be. By the time it runs, the approval refers to something that no longer exists.
THE SINGLE OPERATOR
One engineer has the credentials to do the irreversible thing and no technical reason to pause. The control is that they are careful.
THE MISSING RECORD
Six weeks later, nobody can establish who authorised the change — only that it happened.
WHAT TALOS DOES
Control that survives a bad day.
Multi-stage pipelines
Compose build, test, approve and deploy stages. Each stage is a job with its own logs and its own retry.
Gates that pause mid-run
The pipeline stops at the gate with the pending change shown, rather than asking for permission before anything is known.
Two-person verification
The approver must be a different person, and must independently hold the permission for the action they are approving.
Diff at the gate
What will change is rendered against what is currently running, so approving is an informed act.
Retry without rerun
A failed stage is retryable on its own. A four-stage pipeline does not restart because stage three timed out.
Recorded outcomes
Who proposed, who approved, what ran, and what it produced — one entry, permanently attributable.
HOW IT RUNS
Every action is a tracked job.
WHAT MAKES THIS DIFFERENT
The approval refers to the change, not to the intention.
A gate before the pipeline starts approves a plan. A gate inside the pipeline approves a fact. Talos pauses at the stage itself, with the built artefact and the computed diff already in hand, so the approver is looking at what will actually happen rather than at what someone said they were going to do. The pipeline holds its place until they answer.
RELATED CAPABILITIES
Put a gate in front of one thing.
Start with restores or production deploys. Everything else can stay ungated.