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

ALGAL

Language and VM for long-running agent programs that pause for approval, resume after restarts and replay, with typed decision cells.

Projectresearch
View on GitHub ↗Live demo ↗

Overview

ALGAL is a programming language for agents that run for weeks. A program is data: a typed graph of steps, model calls and tools that a host can inspect, hash and run. It can pause for an approval, exit and resume days later on another machine without redoing work, and every model call and tool result is recorded so runs can be replayed and checked offline. A Lisp-shaped surface and a readable source format compile to the same program digest.

Model judgment is typed: decisions sit in cells with declared context and budgets, and ordinary computation supplies the discipline around them. Programs also evolve by proposing a candidate, evaluating it against the current version on declared cases, and adopting the better one. The demos answer the decision step with a scripted adapter; the list summary places typed decisions (Jev-style) at the branch points.

What the decision model decides

TypeQuestionAnswers
ChoiceDecision cell at a program branch point (e.g. recommend or propose in a release review)(inferred)—

How it works

  1. Write or generate an ALGAL program
  2. Host runs it, calling typed decision cells at branch points
  3. Program pauses for approval and can resume later
  4. Recorded effects allow offline replay and verification

Strengths

  • Resumes without repeating decisions or effects
  • Offline-verifiable receipts
  • Same checks for human- and model-written procedures

Caveats

  • Jev is not named in the README excerpt; demos use a scripted adapter
  • New language with a learning curve

Getting started

bun scripts/vm-demo.ts runs the release-review demo locally.

At a glance

Type
Project
Status
research
Language
TypeScript
Runtime
VM / CLI (Bun)
Has code
yes

Use cases and integrations

How curated lists describe it

  • A language and VM for agent programs that wait for approval, resume after a crash and replay what they did, with typed decisions at the branch points.

    MrJev/awesome-jev

Frequently asked questions

Where do decisions happen?

In typed decision cells with declared context and budgets at program branch points.

What happens after a crash?

The program resumes from recorded effects without repeating completed work.

Is a Jev key needed for the demo?

No, the demos use a scripted adapter for the decision step.

Related projects

  • pi-fabric

    Programmable tool and agent runtime for Pi with one fabric_exec tool; includes an optional Jev loop for observe-decide-act.

    ServicebetaTypeScript
  • SemDecide

    Unix CLI for typed semantic decisions in pipelines and CI: Jev predicates, routes, scores and filters with predictable exit codes.

    CLIbetaPython
  • AutoGPT

    AutoGPT agent platform, listed as adding Jev decision blocks for typed routing, filtering, scoring and confidence-gated next actions.

    ProjectproductionPython
  • Bops

    Mac app with AI bots running business ops on cloud computers, using TypeSafe for small yes/no calls like risk and which bot takes a job.

    Projectproduction
  • opencompany

    AI workspace with chat, durable tasks, integrations and a wiki whose approval review for workspace actions runs through Jev.

    ProjectproductionTypeScript
  • SystemOneHarness★ 193

    Agent harness that compiles an environment's finite actions into typed questions for Jev and gates each step by confidence.

    CLIbetaPython