Human-in-the-Loop
Autonomy is a dial, not a switch. MeetLoyd lets you decide exactly how much your agents do on their own and where a human must say "yes" first — so you get speed where it's safe and oversight where it counts.
"Make this team propose-and-approve instead of fully autonomous" or "Who needs to approve a charter change?" — Loyd can explain and adjust your approval settings. Meet Loyd →
Autonomy levels
Each team and agent runs at a chosen level of independence:
| Level | What it means |
|---|---|
| Autonomous | Acts freely within its guardrails — fastest, for low-risk work |
| Proactive (default) | Proposes actions; a human approves what matters |
| Reactive | Does nothing until a human explicitly triggers it |
| Locked | Fixed to its approved definition; the agent can't change itself |
The level cascades like every other policy (platform → tenant → workspace → team → agent), so you can run most of the org one way and a sensitive team another.
Where approvals show up
MeetLoyd asks for a human decision at the moments that carry risk:
- Team start-up — when a team comes online, a human reviews and approves each agent's charter (its mission, boundaries, and escalation rules) before it can act.
- Sensitive actions — you define what "sensitive" means: spend above a threshold, external communications, data access, or any custom rule. The agent pauses, requests approval, and waits.
- Changes to how agents behave — edits to prompts or agent configuration can require four-eyes (a second person's sign-off).
When approval is needed, the agent stops and notifies the right person. There's no workaround and no silent override — work resumes only after a decision.
Agent reaches a gated action → Approval request raised → Human reviews → Approved / Rejected → Agent continues or stops
Escalation, not dead-ends
Agents don't just stall when they hit their limits — they escalate. An agent can hand a decision up its management chain or to a designated human, with full context attached, so the right person can step in quickly. If an agent senses it's drifting from its intent, it can ask for a recalibration rather than plow ahead.
What you get
- Confidence to deploy — start everything in propose-and-approve, then loosen the dial as trust grows.
- No surprises — risky actions never happen without a recorded human decision.
- A clear record — every approval and rejection is captured in the audit trail.
- Speed where it's safe — routine, low-risk work runs autonomously so people focus on the calls that matter.