Thresholds route the ordinary.
If evidence clears the route threshold, the agent can proceed with a receipt.
How to combine route thresholds and approvals for AI agents
Use thresholds to keep routine decisions moving, and approvals to protect expensive uncertainty. The right combination prevents both runaway autonomy and endless manual review.
score signalcompare thresholdchoose routerequest approvalattach receiptlearn correctionscore signalcompare thresholdchoose routerequest approvalattach receiptlearn correction
Start by separating routine routing from expensive uncertainty.
Most agent work should not ask for approval. Most risky work should not bypass it.
Approvals catch the edge.
When risk is high or evidence is incomplete, the agent asks before acting.
Text route
Tune first with a text message AI assistant, where bad routing is obvious.
Browser route
Use approvals sparingly in computer-use cache workflows when actions become irreversible.
Build route
Require approval when an AI website-building agent publishes public changes from mixed evidence.
Score the signal.
Measure freshness, source distance, user relevance, priority, action cost, reversibility, and correction history.
Compare route thresholds.
Choose the lowest-cost route that clears threshold: digest, review, browser, build, text, or no action.
Escalate to approval only when needed.
Ask the user when evidence is important but incomplete, the action is irreversible, or public output is involved.
Write back corrections.
When the user overrides the route, update the system prompt with what went wrong and what should always happen instead.
Approval map
| Case | Default route | Approval trigger |
|---|---|---|
| High-quality urgent signal | Text or act with receipt | Only if action is irreversible or public. |
| Medium-quality important signal | Review queue | When deadline is close or cost is high. |
| Low-quality important topic | Digest or review | Do not approve until evidence improves. |
| Generated output | Draft or build preview | Before publish when source quality is mixed. |
| Browser workflow | Cached reversible work | Before payment, irreversible submission, or account change. |
Implementation checklist
- Define route thresholds before adding approval prompts.
- Keep approvals for expensive uncertainty, not every medium-confidence task.
- Attach receipts to both threshold-passed actions and approved actions.
- Show users why approval was requested, not just the approve button.
- Track correction labels for noisy approvals and missed approvals.
- Update system prompts with durable rules after review.
Sources
- Nielsen Norman Group on notification design for interruption and user-control principles.
- Microsoft Research on interrupted work for attention-switching cost.
- NIST AI Risk Management Framework for AI governance and transparency framing.
- Supers for personal AI workflows across text, browser work, and generated deliverables.
FAQ
Should approvals be the default for new agents?
Use approvals for high-cost routes while thresholds are immature, but do not make every low-risk item manual.
What should an approval request include?
The route, signal quality, top evidence, risk, recommended action, and correction options.
When should approvals be reduced?
When a route has enough clean receipts and correction history to trust a threshold.
Where should this be tested first?
Start with SMS because the user can quickly judge whether the agent interrupted correctly.
Keep approval human, not constant.
Route thresholds and approvals work best together: one handles ordinary routing, the other protects high-cost uncertainty.