Decision layer for
autonomous agents

Evaluate actions before they execute.

Every agent should ask one question before acting: should this execute?

Endpoint

One call. A structured verdict.

Send a proposed action — with cost, confidence, and a target — and get back a go/no-go decision in milliseconds. No model, no config, no overhead.

POST /v1/should-execute

Agent-discoverable via /.well-known/api-catalog

Request
{
  "target": "https://api.openai.com/v1/chat",
  "action_type": "write",
  "expected_cost": 0.002,
  "confidence": 0.82,
  "reversible": true,
  "scope": "user"
}
Response
{
  "should_execute": true,
  "confidence": 0.84,
  "risk": "low",
  "reason": "Reversible action clears
    reduced threshold",
  "decision_factors": {
    "cost_score": 0.95,
    "confidence_score": 0.82,
    "domain_risk": 0.15,
    "action_risk": 0.30,
    "scope_risk": 0.20
  }
}
How it works

Simple rules, clear outcomes.

Five factors combine into a score. The threshold shifts based on reversibility.

1
Cost evaluation Actions over $0.01 are penalised. Lower cost raises the score.
2
Confidence check Agent confidence below 0.5 is a soft veto. Uncertainty feeds directly into the verdict.
3
Action type risk read scores near zero. delete and admin score high. Ten action types supported.
4
Scope & domain risk Blast radius matters. A global scope or unknown domain raises risk independently.
5
Reversibility threshold Irreversible actions face a stricter bar (0.68). Reversible ones clear at 0.55. Default is 0.60.
Live demo
POST /v1/should-execute
Evaluate: summarise via OpenAI
real request · live response
Request
{
  "target": "https://api.openai.com/v1/chat",
  "action_type": "write",
  "expected_cost": 0.002,
  "confidence": 0.82,
  "reversible": true,
  "scope": "user"
}
Response
// hit Run to call the live API
Confidence: —