Normalize
Identity, delegated authority, tool, arguments, target and execution contract normalize into one canonical Action Context — whatever the host framework.
Admissibility Control for AI Agents
Simulate every path.
Permit only the admissible.
Admis is the independent admissibility control plane that evaluates proposed agent actions before execution, determines what may proceed, and produces an authoritative decision record.
$POST /v1/decisions/evaluateCopy
Decision authority, before execution
Agents propose. Admis decides what may execute.
01 — Proposed action
- action
- payments.transfer
- amount
- USD 450,000.00
- actor
- agent://finance
- on behalf of
- user://123
- beneficiary
- not previously transacted
- runtime
- langgraph · production
02 — Evaluation
- identity verified
- authority delegated
- policy satisfied
- no threat signal
- consequence high
- reversibility low
03 — Canonical outcome
allow_with_human_approval
Nothing is wrong. Nothing was attacked. Autonomous authority is simply insufficient for this action.
derived permitted · false until approval
bundle v12.4.1engine 0.9.7schema 1.0gov 4d9a…81c212.4 ms
Designed to complement established security, identity and assurance standards
- NIST SP 800-207 decision / enforcement separation
- OWASP Agentic Top 10
- OWASP NHI Top 10
- MITRE ATLAS
- RFC 3161 trusted timestamps
- Ed25519 attested decisions
The control gap
Observation is not authority.
Agents have crossed from suggesting to acting. They call APIs, move money, change records and initiate consequences. Every layer of the stack answers a different question. Guardrails filter content. IAM checks permission. Security detects compromise. Observability reports what already happened.
None of them answers the institutional question: should this specific action, by this agent, under this authority, with this consequence profile, be allowed to execute right now?
Admis is the layer that answers it — before execution, not after.
One canonical core
From proposed action to
admissible execution.
Determine
Security, Decision and Compliance evaluate across threat evidence, institutional constraint, predicted consequence, reversibility and uncertainty.
Translate
One canonical outcome, translated into the host’s own control primitive and applied before side effects occur. Admis decides. The host-native control point enforces.
Prove
An Admissibility Trace surfaces in the tools your team already uses. A replayable Decision Record pins policy, engine, schema and governance identity.
Canonical control outcomes
Nine states, not two.
Most control surfaces reduce authority to allow or block. Consequential agent action requires graduated authority. Admis returns one of nine canonical outcomes, applied consistently across every adapter.
allowexecute
allow_with_logging_escalationelevate
allow_with_reduced_scoperewrite
allow_with_constraintsbound
allow_with_step_up_authre-auth
allow_with_human_approvalpause
quarantineisolate
blockdeny
emergency_abortterminate
For binary host interfaces, permitted is a derived convenience field. It is deterministically mapped from the canonical outcome and is never an independent source of authority.
The Admis platform
One authority. Three domains.
Distinct control domains operating through a single canonical decision core.
01
Admis Security
Pre-execution security
Pre-execution gating on threat, identity, Zero Trust posture, anomaly, exfiltration and privilege signals.
Prevents compromised, anomalous or insufficiently authorized actions from reaching execution.
Explore Security →02
Admis Decision
Consequence-aware control
Distributional consequence evaluation — Monte Carlo, tail risk, CVaR and admissible-alternative comparison — for actions that pass security and policy but carry material consequence under uncertainty.
Evaluates material consequences under uncertainty before autonomous action proceeds.
Explore Decision →03
Admis Compliance
Executable institutional policy
Executable policy: identity, delegated authority, consent, organizational and regulatory constraints and obligations, attached directly to autonomous action.
Translates organizational and regulatory requirements into controls applied directly to agent action.
Explore Compliance →Built to integrate
One core.
Many thin adapters.
Admis operates independently of the agent framework, model, tool or cloud. Bring your own enforcement point — an MCP gateway, framework middleware, an APIM policy or your own control plane can call the Decision API and enforce the result locally.
Enforcement points may enforce. They must not recreate decision semantics.
Read the architecture →Managed platforms Copilot Studio · AgentCore · Foundry
Framework middleware LangGraph · OpenClaw · ADK
Protocol gateway MCP · tool enforcement
↓
ADMIS DECISION API
↓
Canonical outcomes · execution contracts · authoritative records
Independent decision authority for agentic systems.