Separate the action.
Challenge the outcome.
Infrastructure between existing agents and existing software. Designed around explicit permissions, scoped claims and independent evidence.
User prompt
Existing user intent.
Existing agent
Your agent remains the agent.
ANL gateway
Typed SDK / HTTP; optional MCP tools.
Control server
Tenant registry, policy and current qualification.
Local worker
Permissions and bounded actions.
Target app
Existing, unmodified software.
Expected effects + forbidden effects + fresh evidence.
Independently implemented checks, not an action pathway confirming itself.
Discover in the sandbox.
Execute in production.
Start with intent
EXISTING PROMPT → EXISTING AGENTThe existing agent receives the user’s prompt. ANL sits beneath it, not in its place. The gateway resolves a scoped business operation, not an unrestricted UI instruction.
Observe in isolation
SANDBOX / DISCOVERYIn an authorized, disposable test environment, the worker encodes permitted DOM and ARIA structure into a typed observation IR. Windows UIA is a future direction, not finished native support.
Experiment safely
SANDBOX / CAUSAL HYPOTHESESBounded experiments induce candidate capabilities from partial, observable business transitions. These are causal hypotheses, not universal decompilation or access to hidden internals.
Synthesize a contract
SANDBOX / FALSIFIABLE SCOPEDescribe typed inputs, permission policy, preconditions, expected effects, forbidden effects and evidence requirements. Every claim must be narrow enough to challenge.
Attempt to falsify
SANDBOX / ADVERSARIAL TESTSChallenge wrong targets, stale observations, false success toasts and forbidden side effects. An independently implemented verifier evaluates the contract rather than trusting the action pathway.
Qualify, then activate
CERTIFICATION / HUMAN DECISIONA scoped capability certificate records qualification and its limitations. Human activation is separate and explicit. Qualified does not automatically mean active.
Publish typed tools
REGISTRY / CURRENT QUALIFICATIONExpose only explicitly activated and currently qualified capabilities as typed SDK / HTTP operations and, optionally, MCP tools. Certification is a capability state, not a promise about every future invocation.
Execute within bounds
PRODUCTION / ACTIVE + CERTIFIED + AUTHORIZEDThe server coordinates an approved intent. The customer-side worker independently enforces local authorization and performs bounded UI actions. No unapproved exploratory production writes.
Challenge fresh readback
PRODUCTION / INVOCATION VERDICTFresh authorized observations are checked for expected and forbidden effects. Return VERIFIED, FAILED or UNKNOWN with evidence tiers and limitations. UI readback alone is not an external backend authority.
Suspend on drift
LIFECYCLE / RECERTIFICATIONChanged interfaces, policies or failed qualification can suspend a capability. Recertification and explicit activation restore eligibility; stale certificates must not permit production execution.
The server coordinates.
The worker enforces.
The verifier challenges.
Dedicated ANL server
Tenant-scoped registry, policies, discovery and capability qualification. Explicit human activation before publication.
Endpoint worker
Encodes permitted interface observations and translates approved intents into bounded UI actions. Enforces local permissions.
Fresh authorized readback. Expected and forbidden effects. Evidence-bound verdict.
VERIFIED / FAILED / UNKNOWNNot all evidence
is equal.
Fresh UI readback
The expected state was read back through a fresh authorized observation. This is visible-interface evidence, not proof of an authoritative backend state.
Local credentials.
Explicit authority.
The initial test bed
A synthetic browser invoice / finance sandbox is the initial MVP. No published benchmarks, completed security audits or certified real-world transactions are claimed.
What ANL is not
Not an agent replacement, universal decompiler, RPA robot or generic MCP generator. It is an in-progress capability qualification and outcome verification layer.