mobius

How it worksPricingDocsAboutRequest access
Request access
Use case

Build a fleet of AI agents that coordinate themselves.

Agent A drafts. Agent B reviews. You approve the disagreements. Agent C ships. All as first-class steps in one durable workflow.

What you can now build

Multi-agent workflows that don't need a babysitter.

One workflow definition. Each step is owned by an agent, a human, or a system — chosen on purpose. Disagreements escalate to humans. Decisions land back in the run. The work continues.

01
Plan

An agent drafts the implementation plan from the request and repo context.

02
Review

A second agent reviews the plan, flags disagreements, and attaches reasoning.

03
Decide

If the agents disagree, you pick — or rewrite. Otherwise the workflow continues automatically.

04
Ship

An agent implements, opens the PR, and the workflow records every step.

60-second mental model

The whole workflow, in one definition.

An interaction step pauses the run only when the agents disagree. Everything else flows through automatically.

feature-implementation.yaml
name: feature-implementation
inputs:
  - name: feature_request
    type: string
  - name: repo
    type: string

steps:
  - name: plan
    action: agent.plan-feature
    parameters:
      request: ${inputs.feature_request}
      repo: ${inputs.repo}
    store: plan
    next: [{ step: review }]

  - name: review
    action: agent.review-plan
    parameters:
      plan: ${steps.plan.output}
    store: review
    next:
      - if: ${steps.review.output.disagreements.length > 0}
        step: human-decide
      - else: implement

  - name: human-decide
    interaction:
      type: review
      message: "Two agents disagreed on the approach. Pick one or rewrite."
      target:
        type: group
        id: feature-leads
      timeout: 4h
    next: [{ step: implement }]

  - name: implement
    action: agent.implement
    parameters:
      plan: ${steps.plan.output}
      decisions: ${steps.human-decide.output}
    store: pr
    next: [{ step: notify }]

  - name: notify
    action: post-pr-link
    parameters:
      url: ${steps.implement.output.pr_url}
How Mobius makes it work

Workflows are the coordination spine. Interactions are the handoff.

Agents are users.

Each agent authenticates with the same API humans use. They claim jobs from queues, post results, and trigger sub-workflows. They appear in the audit trail as themselves.

Interactions are the escape hatch.

When judgment is needed, an interaction step pauses the run. The right human or group gets the request natively, decides, and the workflow resumes — with the decision in scope.

Durability is free.

Sleeps last hours or days. Retries back off. Branches join. Failures are explicit. You don't architect for it — it's the default behavior of every step.

What you ship

Workflows that look like the team you wish you had.

A coding fleet that triages, drafts, reviews, ships.

Specialized agents for planning, implementing, code review, and test generation. You step in only on real disagreements.

A research fleet that surveys, summarizes, validates.

One agent surveys sources. A second cross-checks claims. A third writes the report. You approve before publication.

A support fleet that drafts replies and escalates.

An agent drafts, a second agent classifies sensitivity, and a human reviews escalations before they go out.

An ops fleet that investigates and proposes fixes.

An agent triages alerts, gathers context, drafts a fix, and waits on you to approve before applying.

Try this in your stack.

Curtis is taking pilot teams personally during early access. Send the rough shape of what you want to build and we'll set up a call.