Arxi Decision Gate Asset Core
Docs

Decision Gate Basics

Programs are reliable because requirements are explicit, inputs are known, and outcomes can be verified. Agents are powerful, but their workflows are probabilistic and their results are hard to prove.

Decision Gate recreates those program guarantees for agent workflows without turning agents back into programs. It does it through three concepts that build on each other: Logic defines what “done” means, Evidence proves what happened, and Gates make the decision.

Three Core Concepts

Logic (How Requirements Are Expressed)

Programs encode requirements inside code paths. Agents need requirements stated explicitly, in a form both humans and systems can inspect. Logic turns “did it work?” into clear, testable conditions and combines them into a gate.

{
  "requirement": {
    "And": [
      { "Condition": "tests_ok" },
      { "Condition": "coverage_ok" }
    ]
  }
}

A simple gate requirement: both conditions must be true for the gate to pass.

Why this matters: you stop debating outcomes and start evaluating them. Logic makes success criteria visible, stable, and auditable.

Evidence (Where Proof Comes From)

Programs know their inputs; agents generate results through tools, APIs, and external systems. Evidence is how Decision Gate anchors agent claims to real data.

  • Verified: provider-pulled, trustworthy, slower (use for audited runs)
  • Asserted: client-supplied, fast iteration, lower trust
  • Audit: recorded metadata that does not affect gate outcomes
{
  "evidence": {
    "tests_ok": { "lane": "verified", "value": 0 },
    "coverage_ok": { "lane": "verified", "value": 92 }
  }
}

Evidence payloads and trust lanes: the data the gate will evaluate, plus how trustworthy it is.

Why this matters: evidence turns “the agent said so” into “the data proved it.” It lets you move fast in iteration and still get formal proof when it counts.

Gates (Formal Checkpoints)

Gates are the decision points that turn logic + evidence into outcomes. They evaluate evidence deterministically: same evidence yields the same result. Missing or untrusted evidence becomes unknown, so the gate holds instead of passing.

{
  "gate_id": "quality",
  "result": "pass"
}

The simplest possible outcome: the gate evaluated the evidence and passed.

Why this matters: you get program-grade guarantees—deterministic decisions, fail-closed behavior, and replayable audit trails—inside agent workflows. Each live run produces a runpack so decisions can be verified later without re-running tools.

Next, see how these concepts appear in real workflows in Decision Gate Examples.

Learn More