ReadyForAIDigital workforce infrastructureBook a demo

Security & Permissions

How many reviews a Skill passes before it enters the enterprise, what each Agent can see and do, who intercepts high-risk actions, and why the audit trail lives outside the Agent runtime.

Multi-stage Skill review, role-based isolation, synchronous policy blocking, and an audit trail outside the runtime

  • Skills pass the LarkScout three-layer gate and deploy only after human authorization; HarnessServer SkillScan can add a pre-deploy scan
  • Agents receive different content by Coordinator / Worker role and class; secrets stay out of context, and memory and knowledge interfaces are off by default
  • NodalOS evaluates runtime policies with CEL, and OwlAudit evaluates synchronously on hooks — denying, freezing, or routing to human review on a hit
  • OwlAudit holds the audit ledger independently; written records are verifiable with gap checks, and human approvals happen in AULO
Unvetted Skills go straight to productionEvery Agent sees the same contentAgents call production APIs uncheckedCredentials scattered across Agent codeNo complete audit trail for investigations
  1. 01Upload a Skill package (ZIP / Tar.gz)
  2. 02Archive safety: zip bombs / Zip-Slip / symlink escape
  3. 03SKILL.md and directory checks
  4. 04Process-isolated YARA scan
  5. 05Human confirmation and authorization
  6. 06Optional SkillScan, then deploy

The three-layer ingest gate blocks critical / high matches by default; the YARA layer needs the scanner deployed, and prompt-injection rules do not hard-block by default. SkillScan is off by default and, when enabled, statically scans Python before deployment. These are all static checks, not a runtime sandbox.

Role-Based Visibility

HarnessServer distributes different Skills, SOPs, and Rules by Coordinator / Worker visibility and by Agent class, skeleton, and purchased products

Gated Memory and Knowledge Access

Managed memory is off by default and queryable only by Agents granted memory_access; the LarkScout knowledge interface is off by default and accepts mTLS connections only

Secrets Stay Out of Context

Secrets are managed via File / Env and never enter Agent context

CEL Runtime Policies

NodalOS evaluates runtime policies dynamically with CEL (Common Expression Language)

Synchronous Blocking and Review

As a synchronous evaluator on connected hooks, OwlAudit can deny an action, write a freeze record, or route it to HITL review according to policy

Audit Ledger Outside the Runtime

The audit chain and review tickets are held by OwlAudit, not inside the Agent runtime; written records can be recomputed, verified, and checked for gaps

ProblemProductCapability
Unvetted Skills enter the enterpriseLarkScoutThree-layer ingest gate: archive safety, format checks, and process-isolated YARA scanning (scanner deployment required)
Teams want one more check before a Skill deploysHarnessServerDeploys only after human confirmation and authorization; SkillScan can add Python AST and taint scanning (off by default)
Every Agent sees the same contentHarnessServerDistributes Skills, SOPs, and Rules by role and configuration
Runtime policies are scattered across AgentsNodalOSA CEL policy engine evaluates runtime policies in one place
Agents call production APIs uncheckedOwlAuditPolicy hits trigger the configured action: HITL / HOTL / HOOL intervention, with synchronous deny and freeze
No complete audit trail for investigationsOwlAuditHash Chain audit ledger with recompute-and-verify and gap detection; policy versions increment in the same transaction as audit records

By role

Technology leaders

Static review, role isolation, and runtime policy form separate layers, each with a clear owner

Compliance & risk

The audit chain is independent of the Agent runtime and written records are verifiable; policy changes carry versions, so the policy in force at the time can be traced

Business leaders

High-risk actions pass policy checks and human review first, and there is an audit record to consult afterwards

What else do you want to know?

Private deployment & security boundariesHarness EngineeringMulti-Agent CollaborationBook a demo
0 / 500
SYSTEM READYintent/security-permissions