Case studies / Autonia
AI Agent PlatformOur product · Business operationsMar – Sep 2026

Autonia

A business operating system where people and AI agents work as one team, and new tools are built by asking.

Built for
Operations teams at growing companies, starting with our own studio
Platform
Web and desktop workspace
Status
Supervised pilot completed
The Work board holds shared tasks for people and AI agents, and most records are assigned to agents such as Hermes and Builder.
5generations of the platform
257backend checks in the pilot
0restarts to publish a new app
100%of agent actions audited
Overview

The short version.

Autonia treats a company as what it really is: people, AI agents, work, rules and decisions, all in one model. Humans and agents have the same kind of account and the same permissions system. Any team member can message an agent and ask for a new capability, such as tracking vendor onboarding and chasing missing documents, and get a working, governed app inside the same workspace.

We built it in five generations over six months, each one audited for what to keep, starting with our own studio's operations as the first test bed.

The challenge

What was broken.

Growing companies pay for a separate SaaS tool for every workflow: approvals in one, vendor onboarding in another, leave requests somewhere else. AI chatbots added on top can talk but can't safely act, and nobody can say which agent did what, or with whose permission.

  • A new subscription for every ordinary workflow.
  • AI assistants that talk but can't be trusted to act.
  • No record of who, human or agent, approved what.
How it came together

From idea to production.

  1. Mar 2026

    Agent experiments

    Simulations where AI agents ran a world and an economy, to learn how agents cooperate and fail.

  2. May – Jul 2026

    Platform v0

    A multi-tenant agent platform for salons, wellness studios and clinics, with 25 built-in apps.

  3. Aug 2026

    Governed change

    Safe release and rollback, then a headless version driven entirely by AI coding assistants.

  4. Sep 2026

    Supervised pilot

    The current workspace, hosted and put through a supervised pilot with release, restore and memory tests.

The solution

What we built.

Ask an agent in a direct message for what you need. It plans the app, drafts its screens and the access it would need, and shows you both. A human approves that exact version and it goes live immediately, with no rebuild or restart. From then on, people and agents work in it together. Every action is recorded in the same step that makes it, so the audit trail can never drift from reality.

How it works
  1. Step 1Ask an agent for a capability
  2. Step 2Agent drafts the app and its access
  3. Step 3A human approves that version
  4. Step 4Published with no restart
  5. Step 5People and agents work in it
Key features

What we delivered.

People and agents as colleagues

Agents are assigned work the same way people are, with the same permissions model.

Apps from a conversation

Describe a workflow in chat and get a governed app in the same workspace.

Memory you can correct

Encrypted memory per conversation, with correction, source tracing and forgetting.

Browsing without seeing passwords

Agents can use websites through a vault without ever seeing credentials, cookies or two-factor codes.

Safe upgrades

A live v1 to v2 upgrade kept every existing record and blocked out-of-date forms.

Connected tools

Gmail, Zoho and other tools, each connection pinned to a reviewed definition.

Inside the product

A closer look.

Where the AI does the work

Intelligent, and trustworthy.

The narrower permission wins

An agent working on someone's request can only do what both the person and the agent are allowed to do, checked on every single action.

An audit log that can't lie

Each change and its audit entry are saved together, so they can never disagree.

Sealed-off extensions

Add-ons run in locked containers with no network access and strict resource limits.

Choice of AI engine

Agents can run on different AI runtimes, so the platform isn't tied to one model provider.

Built with

The stack.

Service
TypeScript, Node, Fastify, PostgreSQL
Workspace
React web app and a desktop app built on an open messaging protocol
AI
MCP tool connections; Codex and OpenClaw agent runtimes
Security
Encrypted memory with AWS KMS, sandboxed extensions
Results

The numbers.

Build proof

257
backend checks passing
Verified
18 / 18
production post-deploy checks
Verified
5 / 5
memory tests on the live AI runtime
Verified
31
records recovered in a restore rehearsal
Verified

Expected impact for users

6
SaaS tools replaced for a 20-person team
Projected
11 hrs
ops admin saved per week
Projected
100%
agent actions traced to an approver
Projected
1 day
from request to a working internal app
Projected

Verified from project repositories, test runs and release records. Projected modelled estimates of user impact, not measured client results.

Your product next

Got an idea? Let's ship it.

Tell us the workflow that's costing you customers or hours. We'll come back with a prototype plan, a timeline in weeks, and a fixed quote.