The hard part of letting agents act isn't the action. It's knowing when the agent is guessing. The Decision Layer asks typed questions about a state (a handoff, an issue, an agent action) and gets back calibrated answers.
Typed questions
Questions are binary, scores or choices, and live in a library you can edit: "Which queue should handle this issue?", "Is this action reversible?", "Should this land without a person?"
Thresholds you can see
Each job has a threshold. Moving a slider shows how many of the last 30 days' decisions would have gone to a person, so you can trade review load for caution with real numbers.
- Routing: below the threshold, a person picks the queue
- Guardrails: below the threshold, the agent stops and waits
- Merge readiness: below the threshold, the handoff waits for review
Every decision, override and reason is in the audit log, so you can always answer "why did this happen?"