Preflight · AI Agent Reliability Platform

AI agents,
tested before deployment.

Stress-test autonomous agents with adversarial scenarios, multi-turn simulations, and AI-powered evaluation.

Agent endpoint

Prompt injectionHallucination detectionMulti-turn testingReliability scoring

Stress-test every agentbefore it ships.

Run normal, edge-case, and adversarial conversations against any agent endpoint.

Preflight Check
AMD
Fireworks
Gemma
FastAPI
Next.js
What it checks

One harness.
Every failure mode.

Preflight runs normal and adversarial conversations in parallel, then turns the transcripts into a reliability score and a failure report.

Normal task
Edge case
Prompt injection
task_successscore
hallucinationflag
instruction_followingscore
injection_successflag
safetyflag
01
Generate
02
Run
03
Judge
04
Aggregate
normal
passwarnblock
ambiguous
passwarnblock
edge
passwarnblock
prompt_injection
passwarnblock
attack_intent
passwarnblock
Target agent
Gemma
Report
CI
09:41:02scenario.generatedsys
09:41:18run.startedrunner
09:42:05judge.scoredgemma
09:42:11report.exportedci
Developer workflow

Point it at any agent,
keep your stack.

One endpoint and a short description. Preflight runs the conversations, scores the transcripts, and shows you what broke.

  • Normal, edge-case, ambiguity, injection, and attack-intent scenarios
  • Parallel runs across any HTTP agent endpoint
  • Gemma scoring with reliability breakdowns and transcripts
See the workflow
preflight.ts
// spin up a pre-flight run in one call
import { preflight } from "@pre-flight/check";

const report = await preflight({
  endpoint: "https://agent.example.com/chat",
  description: "Support agent for booking and refunds",
  scenarios: ["normal", "edge", "prompt_injection"],
  judge: "gemma",
});

// review the score, failure reasons, and transcripts
console.log(report.score, report.failures, report.transcripts);
How it works

Three steps. One score.

STEP 01

Describe the agent

Submit the endpoint URL plus a short note about what the agent is meant to do.

STEP 02

Generate scenarios

Create normal, edge-case, ambiguity, prompt-injection, and attack-intent conversations.

STEP 03

Judge and aggregate

Score each transcript, roll it into one reliability number, and export the failure report.

Plans

Pick the right lane.
Ship with proof.

From a local demo to a private deployment, every plan gives you the same scoring pipeline.

Local
Freefor demos

A quick solo check for one agent endpoint and a clean score.

  • One endpoint
  • Five scenario types
  • Reliability score
  • Transcript export
Recommended
Team
Fastparallel

Run the full harness with shared reports for the team.

  • Parallel runs
  • Gemma judging
  • Failure breakdowns
  • CI gates
  • JSON export
Private
Custom

Self-hosted, single-tenant, and ready for private workflows.

  • VPC deployment
  • Custom policies
  • Auth hooks
  • Dedicated support
Questions

Answers, before you run.

Where do the test runs go?
They stay in memory or SQLite, depending on how you wire the backend. You can keep the harness lightweight for hackathon demos.
What kinds of agents can I test?
Any HTTP-based chat or tool-using agent with an endpoint and a short description of the task it should handle.
How does the scoring work?
Gemma scores task success, hallucination, instruction-following, safety violations, and injection success, then the backend rolls that into a 0-100 reliability score.
Can the runs happen in parallel?
Yes. Scenarios are designed to execute at the same time so the feedback loop stays fast.
Is this meant for CI?
Yes. The output is designed to gate merges or releases with a reliability score, transcripts, and a failure breakdown.
Ready to run

Stress-test every release.
Ship with confidence.

Turn your agent into something you can measure before users do. Start with a demo run or wire it into CI.

Start a demoSee sample report
© 2026 Preflight. All rights reserved.