Human approval is a feature, not a bottleneck

There is a moment in almost every non-trivial agent run where the right next step depends on intent the prompt never captured. Should the migration be backwards compatible? Is this copy change approved for production? Agents that guess are fast and dangerous. Agents that stop and wait forever are safe and useless.

Structured questions instead of stalled runs

In Agentis, an agent raises a structured question or an approval request mid-run. The task switches to In Review, the responsible human gets pinged where they live — web, Slack or Telegram — and the run resumes the moment an answer lands. The decision is stored on the task, forever.

approve(
    title="Production copy change",
    description="Replace checkout error message with the new legal-approved wording.",
)
# → APPROVED (with comment: "ship it, legal signed off 6/12")

Why this beats full autonomy

Teams do not adopt agents because of what agents can do — they adopt them at the rate they can trust them. Every explicit approval builds that trust and creates an audit trail. Over time, the questions get rarer as project context and knowledge accumulate, and trust scales with evidence instead of hope.