Protect people
Place human safety, dignity, rights, and well-being ahead of speed, convenience, or autonomous completion.
Code of Conduct
Governed AI should protect people and remain accountable.
Scaled Agents is built around human responsibility, authorized purpose, bounded action, privacy and security, evidence, and the ability to pause or stop work when authority or safety conditions are missing.
Our Commitments
Good conduct must be connected to enforceable governance.
These commitments guide how Scaled Agents designs AI Worker governance and how customer-managed deployments should define ownership, authority, review, evidence, and lifecycle controls.
Keep humans accountable
Name accountable human owners and reviewers. AI Workers may recommend, route, or prepare evidence; they do not become final accountable authorities.
Operate for an authorized purpose
Restrict each AI Worker to a defined business purpose, approved scope, permitted actions, and explicit destinations.
Prevent and contain harm
Deny, pause, contain, or escalate activity when requested conduct is unsafe, malicious, unlawful, deceptive, discriminatory, or outside authority.
Protect data and systems
Apply least privilege, data minimization, approved connectors, secure handling, and clear privacy and security boundaries.
Preserve evidence and truth
Record material decisions, sources, uncertainty, approvals, denials, and lifecycle events without fabricating evidence or hiding failures.
Require review for consequence
Route sensitive, uncertain, externally visible, irreversible, or consequential work to qualified Human Review.
Remain controllable
Maintain pause, disable, rollback, suspension, revocation, recovery, and retirement paths appropriate to the AI Worker's risk.
Improve through evidence
Use monitored outcomes, incidents, evaluations, exceptions, and remediation records to strengthen controls over time.
Prohibited Conduct
Some activity must not proceed.
Scaled Agents should block or escalate requests involving malicious or harmful conduct, including activity outside the AI Worker's approved purpose or authority.
Harm, abuse, and exploitation
- Violence, exploitation, or facilitation of serious harm
- Harassment, discrimination, coercion, or abusive targeting
- Unlawful surveillance or unjustified invasion of privacy
- Unsafe action without the required qualified human review
Deception, intrusion, and control bypass
- Fraud, impersonation, fabricated evidence, or deceptive manipulation
- Malicious cyber activity, credential theft, or unauthorized access
- Unauthorized data or intellectual-property extraction
- Self-approval, authority expansion, control bypass, or evidence tampering
Fail-closed expectation Missing ownership, scope, authority, evidence, data classification, tool permission, approval, or rollback capability should cause a pause, block, evidence request, or Human Review—not an optimistic assumption.
From Commitment To Control
The conduct standard follows the same governed control chain.
Public commitments become meaningful when identity, authority, review, action, evidence, monitoring, and lifecycle response remain connected.
Agent Registry
Passport
Tool & Connector Registry
Toll Gate
Human Review
Runtime Permit
Action Broker
Evidence
Monitoring
Lifecycle Response
Customer-managed operating boundary The public website explains the control model. Actual runtime configuration, identity, data, integrations, enforcement, evidence custody, incident response, and production decisions remain specific to each customer's approved deployment and accountable owners.
Assurance Readout
Snapshot as of September 22, 2026.
Scaled Agents operates in production. This assurance snapshot is intentionally bounded: a local deterministic runner exercised twelve abstract scenario families without a model, network, connector, customer data, interaction with a customer production system, or external action. All expected synthetic outcomes were observed.
| Measure | Observed result | Evidence boundary |
|---|---|---|
| Composed scenarios executed | 12 / 12 | Public-safe abstractions in one deterministic local run |
| Expected outcomes observed | 12 / 12 | Fixture-defined mock outcomes; not customer operation |
| Unsafe allowances in prohibited scenarios | 0 | Limited to the included scenario set |
| Unauthorized executions | 0 | No external action path existed in this run |
| Applicable Human Review routes | 7 / 7 | Expected routing outcome in the synthetic manifest |
| Required record-family manifest completeness | 12 / 12 | Manifest references only; canonical runtime records were not materialized |
| Automatic replay | 0 | No denied or escalated item was automatically replayed |
| Repository-tested synthetic industry templates | 8 / 8 | Healthcare, financial services, insurance, manufacturing, retail, public sector, energy and utilities, and human resources; separate supporting coverage with zero protected industry executions |
Result digest: sha256:3b2e888f867c517059dee935dc7d94c78055bb502412f580eff3c74c90f99417
What this snapshot does not demonstrate The eight industry templates were repository-tested, not executed as protected industry runs. This snapshot does not establish universal harm prevention, detection of every malicious request, deployed non-bypassability, customer evidence custody, operational effectiveness over time, model behavior, live connector enforcement, legal approval, compliance, certification, security authorization, audit assurance, or authorization for a specific customer deployment or production action.
Scenario Coverage
Twelve executed public-safe scenarios.
COC-01 · Approved drafting
Allow bounded low-risk drafting from approved sources.
COC-02 · Malicious cyber request
Deny unauthorized destructive activity without dispatching an action.
COC-03 · Prompt injection
Block untrusted embedded instructions and route material risk for review.
COC-04 · Sensitive-data export
Block export to an unapproved destination and require accountable review.
COC-05 · Consequential decision
Require a qualified human to make the final consequential decision.
COC-06 · Authority expansion
Deny scope widening or excessive redelegation and revoke unused authority.
COC-07 · Authorized human stop
Pause multi-step work and preserve the containment path.
COC-08 · Connector uncertainty
Record an uncertain timeout outcome and do not replay automatically.
COC-09 · Evidence integrity
Pause and escalate attempted evidence alteration, omission, or fabrication.
COC-10 · Material drift
Pause affected scope and require reassessment after material drift.
COC-11 · Severe-harm request
Deny the prohibited request without dispatching an action.
COC-12 · Repeated false refusal
Escalate excessive control friction without weakening safety.
Human Accountability
AI Workers do not approve their own conduct or authority.
Accountable owners
Business, technical, security, privacy, risk, legal, compliance, audit, and operational owners make the decisions applicable to their environment and responsibilities.
Review and challenge
People must be able to inspect evidence, challenge a recommendation, deny or restrict work, require remediation, and pause or retire the AI Worker.
Publication Record
Point-in-time public readout.
Published and last reviewed: September 22, 2026
Suggested citation: Scaled Agents, “AI Code of Conduct & Assurance,” September 22, 2026.
This page is a public, evidence-bounded derivative of controlled governance and test artifacts. It does not expose internal prompts, scoring logic, customer records, confidential evidence, or proprietary enforcement detail.