An honest account of what this is

ApexCorp is a bet that a group of AI agents is more useful organised like a company — with roles, approval chains and escalation — than orchestrated like a script.

Why organise agents at all

A single agent with a long prompt hits two walls. It has no way to keep what it learned separate from what everyone needs to know, so context grows until it is mostly noise. And it has no notion of authority, so there is nobody to escalate to and no second pair of eyes before a mistaken conclusion becomes a shared assumption.

Giving agents an org solves both, but only if the org is real. If “team leader” is just a word in a prompt, an agent will route around it the first time doing so is convenient. So the rules that matter here live in a domain model with tests on them: who may approve what, who may write to which memory pool, who may send anything outward.

What it is honestly at today

• The product works. You can found a company, give it work, watch agents route and review it, and approve what needs you. The demo is that product, running.
• You cannot buy it yet. There is no billing. The pricing page describes what the service would cost to run, not a plan you can subscribe to today.
• There is no registered entity yet. Which is why the legal pages are marked as drafts with bracketed placeholders instead of pretending otherwise.
• It has not been independently reviewed. The safety properties are enforced in code and covered by an automated test suite. That is evidence, not an audit.

How it is built

A pure, immutable domain core that reads no clock and generates no identifiers — which is why its tests need no mocks. An orchestrator holding routing, model policy, safety rails and memory. A Next.js control panel. Dependencies point one way only.

The most instructive bug so far: every QA review was being sent to the top-tier model because the generated review brief tripped a length heuristic, making a one-line typo fix cost $0.93. It was invisible until a real key was used, because a bug that only costs money does not fail any test. The fix took it to $0.0164.

Working with us

We are early enough that the first customers shape the product. If you have a real workload in mind — a support queue, a documentation backlog, a research pipeline — that conversation is more useful to us than a feature request, and usually more useful to you than a sales call.