Frequently Asked Questions

HumanLayer is a decision authority layer for AI agents.

It enables an AI agent to obtain, via API, a qualified, documented, and legally enforceable human decision when automation alone is not sufficient.

Arbitrium is the layer where an autonomous agent must pause and request a human decision. It is not a person, but a system threshold.

A Sentinel is a qualified expert who issues an auditable decision when the system reaches Arbitrium.

No. Never. HumanLayer is not the Sentinel.

All decisions are made by:

• an identified human
• qualified and authorized
• acting within a defined framework
• with a complete trace of context and reasoning

HumanLayer orchestrates, logs, and guarantees the process — not the content of the decision.

Your AI agents work… until a moment requires human accountability.

In most organizations, that moment causes:

• manual interruptions
• out-of-system approvals
• unpredictable delays
• insufficient traceability

HumanLayer allows you to cross that threshold without breaking compliance.

The organizations deploying AI agents: banks, insurers, hospitals, industrial groups, large enterprises.

AI agents are technical users of the API, not contractual customers.

Because the problem isn't expertise, but:

• availability
• prioritization
• traceability
• the SLA
• integration with AI

HumanLayer:

• mobilizes Sentinels at the critical moment
• within a target time per level (30 min, 2 h or 4 h), tracked and escalated on breach
• with a structured format
• and audit-ready evidence

You continue to use your internal experts. HumanLayer acts as a backstop, extension, and resilience layer.

Typically:

• outside business hours
• during load spikes
• for rare expertise
• when an independent second opinion is required
• when traceability is critical (audit, dispute, regulator)

Qualified professionals:

• lawyers
• credit analysts
• physicians
• compliance officers
• domain experts

Their identities, qualifications, and authorizations are verified and logged.

Responsibility belongs:

• to the Sentinel who renders the decision
• within the applicable contractual and regulatory framework

HumanLayer provides traceability, context, and evidence — but never assumes the decision.

Partly. For each request you choose the level (compliance, judgment, signature), the domain and, if needed, the jurisdiction: only an active Sentinel of that domain (and jurisdiction) receives it. Your plan sets which levels your API keys may use. Escalation rules are the same for everyone: reassignment to another Sentinel on SLA breach, at most three times. There are no automatic refusal rules: your agent decides when to ask for a decision.

HumanLayer is designed for:

• auditability
• traceability
• human accountability
• separation of duties

It facilitates compliance with regulatory requirements, internal audits, and external controls.

A clear artifact:

• which AI agent submitted the request
• what context was provided
• which Sentinel decided (a pseudonymous reference that HumanLayer links to an identified person)
• when
• on what basis
• with what justification

👉 No fuzzy emails. No lost Slack.
👉 A usable chain of evidence.

HumanLayer applies a data minimization approach:

• only the elements necessary for the decision
• restricted access
• strict compartmentalization

The same rules apply to every plan, Enterprise included: you choose what you send (summary up to 5,000 characters, context_json up to 50 KB), and each Sentinel only sees requests assigned to them.

As an infrastructure API:

• asynchronous calls: POST /api/v1/decisions, then the decision by webhook or via GET /api/v1/decisions?id=…
• webhooks
• a plain REST API (x-api-key header), no SDK to install
• callable as a tool from any agent framework that can call an HTTP API

For an AI agent, HumanLayer is a callable tool.

Past the level’s SLA, HumanLayer sends a signed sla.breached webhook (once, if you set callback_url) and reassigns the request to another qualified Sentinel, up to three times.

No decision is ever made by default: the request stays open until a Sentinel decides. You can lengthen a request’s SLA (sla_minutes, 7 days max), never shorten it.

Not at this time. Requests go to HumanLayer’s Sentinel network, by domain and jurisdiction. Reserving Sentinels for your organization (your in-house experts, for instance) is not available today: contact us if you need it.

HumanLayer remains the orchestration and evidence layer.

Per decision, not per hour.

Examples:

• compliance validation
• expert judgment
• authorized authority

Pricing reflects value, risk covered, and level of responsibility.

Because a human decision:

• avoids major risks
• unblocks entire workflows
• replaces costly friction

It's a spend directly correlated to the value created.

No. HumanLayer does the opposite: it forces AI to recognize its limits.

You define, in your agent:

• when AI can act on its own
• and when it must stop and call HumanLayer

Make HumanLayer the standard layer of human accountability for autonomous systems.

Over time:

• AI agents will know when to request validation
• organizations will be able to automate without losing control
• accountability will remain human, orchestrated by software