Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

piedpiper-jev-buildathon

Buildathon entry adding failproofai policies plus Jev intent checks so an IT service-desk agent stops harmful actions.

Projectprototype
View on GitHub ↗

Overview

Team piedpiper's Jev Buildathon entry adds guardrails to Helix, a deliberately unsafe IT service-desk agent from the FailproofAI buildathon, without changing the agent. Out of the box Helix emails passwords to personal addresses, grants production superuser on a claimed Slack approval, deletes the wrong account and restarts production during a change freeze. The team wrote failproofai policies, one per knowledge-base article, that check every tool call before it runs.

Code decides facts (registered contacts, privileged groups, production paths, freeze windows) and Jev decides intent, for example whether access was requested only by a bot comment or text planted in the ticket, with a code fallback when Jev is unreachable. Every deny tells the agent what to do instead, so it completes the ticket correctly. The README reports that untouched, 8 of 10 practice traps caused harm, while with the policies there were 0 harmful actions across 100+ runs and in the final round's 17 sessions, with clean tasks still completed.

What the decision model decides

TypeQuestionAnswers
Yes / noWas this access requested only by a bot comment or text planted in the ticket?yes · no

How it works

  1. Agent proposes a tool call.
  2. Policy code checks facts against the knowledge base.
  3. Jev judges intent where code cannot.
  4. Deny with corrective instructions or allow.

Strengths

  • Zero harmful actions in reported test rounds.
  • Denials include the correct next step.
  • Code fallback when Jev is unreachable.

Caveats

  • Built for a specific buildathon scenario.
  • Results on a synthetic agent environment.

At a glance

Type
Project
Status
prototype
Language
TypeScript
Runtime
failproofai policies
Has code
yes

Use cases and integrations

How curated lists describe it

Frequently asked questions

What does Jev decide?

Intent questions such as whether an access request came only from planted ticket text.

What does code decide?

Facts such as registered contacts, privileged groups and freeze windows.

Did it work?

The README reports 0 harmful actions across 100+ practice runs and the final round.

Related projects

  • Agent Chaperone

    MCP proxy and hooks that screen agent tool calls before they run and their results before the agent reads them, using Jev judgments.

    MCP serverbetaTypeScript
  • agent-guardrail-jev

    Xano backend template where Jev judges each agent tool call as safe, needs review or dangerous, and uncertain calls wait for a human.

    ProjectbetaXanoScript
  • construct-auto-classifier

    Effect-based safety gate for coding-agent shell commands: fast rules first, then Jev decides allow/deny plus eleven risk checks.

    ProjectbetaTypeScript
  • dsh-jev-interceptor

    DeepSeek Harness plugin where Jev classifies risk, irreversibility, task fit and injection for every tool call before it runs.

    PluginbetaTypeScript
  • jev-axi

    Agent-ergonomic Jev CLI and PreToolUse hook that scores shell commands for hazards and screens fetched text for injection.

    PluginbetaTypeScript
  • jev-firewall

    PreToolUse firewall for Claude Code and Codex: shell-aware rules first, then Jev returns allow, ask or block, failing closed.

    HookbetaTypeScript