AI Research
What we are learning about putting AI agents to work, and keeping people in charge. Get new pieces by email
OutlineThe approval gate: a hand on the send buttonThe more an agent can do, the more it matters where it stops. We make the case for one hard rule: an agent may prepare anything, but nothing leaves without your yes.Control
OutlineFrom tools to teammates: when AI holds a seatA tool waits to be picked up, while a seat has a job, a manager and a record of its work. We look at what changes when you stop prompting AI and start staffing it.Agents
OutlineZero dependencies: an agent engine you can read in an afternoonThe engine behind the Office uses no third-party packages. Here is why we made that trade, what it cost us, and what it buys the people who run it on their own machines.Engineering
Coming soonMemory that earns its keep: saved choices, not old chatsMost agent memory is a pile of old conversations. We argue for something smaller: when you answer a decision once, it becomes a preference every agent reads before it works.AgentsGet it by email when it is out
Coming soonHow do you measure an AI seat?Success rate, first-try rate and review score, each over enough work to mean something. A look at the bars a seat has to meet before it earns more trust, and why each number is there.IdeasGet it by email when it is out
Coming soonShow the work before it shipsPeople trust what they can check. Why every run in the Office leaves a visible trail of steps, and why the last step before anything goes out is always yours.ControlGet it by email when it is out
Coming soonThe one-person company, staffedWhat a founder can run alone changes when every seat on the org chart can be filled. We sketch how one person, a handful of seats and one Inbox divide the work, and where the person stays essential.WorkGet it by email when it is out
Coming soonFiles as the database: plain JSON you can openApprovals, runs and heartbeats live in plain files on your machine. Why we chose files you can read and back up over a database you have to trust, and where that choice will stop scaling.EngineeringGet it by email when it is out
Coming soonWhat does a career look like for an AI seat?If a seat can earn trust, it can also earn more responsibility. Eight Journey stages, each earned from reviewed work, and what we think promotion should mean for software that works for you.IdeasGet it by email when it is out
Read the brief and your preferences
Waiting for your approvalNothing matches. Choose All or try another word.
Build this with us
Manelink AI makes the Office, where AI agents take seats on your org chart and you see, direct and approve their work.
Meet the company