Skip to content

Platform · Decide

Evaluate every consequential action

Before an agent’s action executes, GREENLIGHT works out what it is really trying to do and decides — deterministically, with the reasons and the deciding policy attached. The same inputs always produce the same decision.

Part of GREENLIGHT

Why it matters

The problem Decide solves

An agent’s request can be ambiguous, manipulated or simply new. A decision layer has to understand what is being asked, apply the rules the same way every time, and refuse to guess when nothing matches. Decide is where every consequential action gets its answer.

Questions it answers

  • What exactly is this agent trying to do?
  • Which rule applies, and what is the outcome?
  • What happens when no rule recognises the action?

Where it sits

Decide in the GREENLIGHT loop

Every consequential action moves through the same loop. Evidence from each decision informs the next policy.

How it works

Decide, step by step

  1. 01

    Describe the action

    Every request is turned into one consistent description: action, target, amount and on whose behalf.

  2. 02

    Scan for risk signals

    Signs of manipulation or unusual behaviour are flagged before any rule runs.

  3. 03

    Apply the rules

    Policies are evaluated deterministically, so the same request always gets the same answer.

  4. 04

    Return the outcome

    Allow, require human approval, escalate or block, with the reasons and the deciding rule.

What’s included

The 4 modules of Decide

Decide · 01

Intent & risk scanning

Each proposed action is described consistently — what it does, to which system, for how much and on whose behalf — and scanned for risk signals before any rule runs.

  • One consistent description for every agent and tool
  • Risk signals detected before evaluation
  • Optional language-model assistance for unclear requests — advisory only, it never decides

Decide · 02

Decision engine

Every action receives one of four outcomes: allow, require human approval, escalate or block — with the reasons, the deciding policy and how well the action was covered.

  • Deterministic and reproducible from the record
  • Anything the rules do not recognise escalates
  • Reasons a reviewer and an auditor can read

Decide · 03

Read safety

Reads are observed rather than blocked by default, and screened for bulk extraction, sensitive data and outbound movement.

  • Writes governed, reads observed
  • Bulk, sensitive and outbound reads flagged
  • Per-connector read handling

Decide · 04

Rollout modes

Start in shadow — Crelis records what it would decide and blocks nothing — then move to advisory and enforce when the evidence supports it.

  • Shadow: observe only
  • Advisory: recommend
  • Enforce: hold or block

In practice

A request no rule recognises

A disputes agent proposes reversing a decision made last month. No rule covers reversals, so instead of letting it through, GREENLIGHT escalates it to a person and records why.

Illustrative scenario.

Use case

Disputes & chargebacks

Who uses it

  • Agents get an answer before they act
  • Reviewers see the reasons behind each outcome
  • Auditors can reproduce any decision from the record
  • Consistent answers

    The same request gets the same outcome, whichever agent asks.

  • Unknowns are not allowed silently

    Anything the rules do not recognise goes to a person.

  • Safe to start

    Shadow mode lets you see every decision before anything is blocked.

See it working

See Decide on your own workflow.

We connect one agent workflow in shadow mode and walk you through every decision GREENLIGHT would make, so you can judge it on your own traffic.