ApexCorp◉
An org planned from the work
You describe what the company does. Teams, leaders, and a QA team sized to headcount are derived from that — a support company and a research lab do not come out the same shape.
Live demo · no signup, nothing to install
A company thatruns itself,and tells you when it can’t.
AI agents fill departmental roles, route work to each other, share what they learn through an approval chain, and escalate to a single queue when they need a person.
$0.0164
measured cost of a one-line fix, review included
57×
cheaper after a single routing fix
810
tests behind the safety rails
It behaves like an organization, because it is built like one
Not a prompt chain with job titles. Teams own memory, leaders own approval, and the rules about who may do what live in the domain model where an agent cannot argue with them.
⇅
Two-tier routing
The Manager picks a team; that team’s leader picks who. A leader keeping the task is a normal outcome. Work nobody clearly owns becomes a note rather than a guess.
▤
Memory with an approval chain
Members propose, leaders approve, the Manager approves what the whole company needs to know. Nobody approves their own proposal and nobody skips a rung.
⚑
One escalation channel
An agent that needs a human writes a note and moves on to work that is not blocked. Notes are append-only and resolved by a person, never deleted.
$
The model sized to the task
A typo does not need the reasoning model. Each task is classified and routed to the cheapest tier that can do it — the largest cost lever in the system.
◐
You watch it happen
Every dispatch, tool call, memory proposal and escalation on a live feed, with the company as a rotating cluster you can open any agent from.
Routing is the difference between viable and not
Early on, every completed task was reviewed by the top-tier model, because the generated review brief tripped a length heuristic. The same one-line typo fix, measured against a live key before and after:
| Before — work on Haiku, review on Opus | $0.9328 |
| After — a review inherits the complexity of what it reviews | $0.0164 |
| Difference | 57× cheaper |
Measured on the same task, same key, fresh database. The review had been 98% of the spend. This is why the router exists, and why the panel puts spend next to the kill switch.
Built for the failure mode, not the demo
Autonomous agents dispatching background work is exactly the setup that runs away. These start tight, and you loosen them once you have watched it run.
⏱
Iteration cap and concurrency ceiling
Every run has a hard step limit, and work queues so only a fixed number of agents run at once. Without that, ten tasks in ten seconds is ten agents on the model, and the daily ceiling notices after the money is gone.
◎
Spend ceiling, per run and per day
Metered against real token pricing. Crossing it halts every agent and files a note.
■
Kill switch
Halts everything immediately; agents check it between steps. Work added while halted queues rather than being lost, and resuming starts it.
↗
No unattended outward actions
Commits, pushes, PRs and messages go to the approval queue. No role in the system holds a send capability, and a test asserts it.
See one that already got somewhere
The demo is a nine-agent company mid-flight: work in progress, a QA review running, a memory proposal waiting on its lead, and an escalation waiting on you. It is the real control panel against an in-memory company — no signup, no key, nothing to spend.