Skip to content
Max Control
Join the list

AI controllers for engineering outcomes

Put entropyon a leash.

Maxwells are persistent software controllers. They observe your systems, act on opportunities, verify the outcome, and keep working until reality matches the desired state.

Join the design partner listRead the controller thesis(opens in a new tab)

The task ends. The objective persists.

Current state

current ≠ desired

Maxwell controller

Desired state

reconciled

Human authority

policy · guardrails · final say

  1. Observe
  2. Compare
  3. Act
  4. Verify
  5. Remember
Control diagram: a disordered current state flows through a Maxwell controller and emerges as an ordered, reconciled desired state. A human authority branch feeds policy, guardrails, and final say into the controller. The controller runs a continuous loop: observe, compare, act, verify, remember.

Early controller results

Internal controller runs. Figures subject to revision.

PRs closed
719
Reverse dependencies
430 → 13
Controller collaboration
4-pass
Production issue detected and resolved
Resolved

Category

Objectives, not tickets.

Agents are usually handed a bounded task. They produce a result, and then they stop. The condition that created the task is left for someone to notice again.

Controllers own an ongoing condition. They keep comparing the system to what it should be, and they keep reconciling the difference.

Agent

  1. Receives task
  2. Produces result
  3. Stops

Controller

  1. Observes state
  2. Reconciles condition
  3. Verifies
  4. Continues

A Maxwell does not wait for someone to rediscover the same class of problem.

Controller catalog / early specimens

Meet the Maxwells

  1. Seismograph

    / Dependencies

    Reduces dependency blast radius by finding overly connected packages and separating them safely.

    Scope
    package graph
    Status
    reconciling

  2. Catalyst

    / Performance

    Finds slow tests and slow paths, makes them faster, and verifies the improvement.

    Scope
    test suite · hot paths
    Status
    reconciling

  3. Surgeon

    / Production health

    Detects production health problems, applies a bounded intervention, and watches the result.

    Scope
    runtime signals
    Status
    watching

  4. Gate E2E

    / Architecture

    Applies senior architectural judgment during review and follows the change through resolution.

    Scope
    pull requests
    Status
    reviewing

Operating sequence

Declare the condition. Let the controller own it.

  1. Install

    Connect the GitHub app or deploy Max Control in your own environment.

  2. Declare

    Define controllers and desired conditions alongside the codebase.

  3. Reconcile

    Maxwells continuously observe, act, test, review, and verify.

  4. Escalate

    When judgment or business authority is required, the system brings a human into the loop and remembers the decision.

Why now / durable control plane

The work can sleep. The context cannot disappear.

A controller is only as good as its memory. Max Control is built to run for months, not minutes: work pauses and resumes, state survives restarts, and every decision a human makes stays on the record.

Max Control

CONTROL PLANE / PREVIEW 0.1

Deployment targets

  • Max Control infrastructure
  • Customer cloud
  • On-premises
Durable execution
Resumable sandboxes pick up exactly where the work left off.
Persistent state
Machine and conversation state outlive any single run.
Explicit policy
Rollback and cost policies are declared, not implied.
Controller collaboration
Maxwells hand work to one another and review each other.
Retained escalation
Human decisions are captured once and applied again.
Harness-agnostic
Runs across multiple coding harnesses rather than one.

Design partner program / limited

What should your team never have to fix twice?

We are working with a small number of engineering teams to define the first controller catalog.