Overview

Havenware runs a fleet of role-specialized AI agents under human governance. The agents do operational work: triage and drafting in your inbox, lead research and sequencing in outreach, post drafting and scheduling on social, invoicing and reconciliation prep in finance, monitoring and document Q&A in research. The governance layer decides what any of them is allowed to do with that work.

The loop

Everything consequential in Havenware moves through one loop: propose, review, approve, execute, audit.

When an agent wants to take an action with consequences outside its own workspace, it does not take the action. It files a proposal in your approval queue: one decidable question, the context needed to answer it, and an honest estimate of how long the answer will take you. The proposal waits. You review it and answer. Only then does the action execute, and the proposal, your decision, and the resulting action are recorded in the audit trail.

This is enforced in code, not policy. There is no prompt asking agents to please check with you first. The execution path for consequential actions requires an approval record; without one, there is nothing to run.

What counts as consequential

Anything external, irreversible, or touching shared state: sending an email, contacting a lead, publishing a post, issuing an invoice, moving money, changing another agent's context. Internal work, drafting, reading, summarizing, organizing an agent's own workspace, does not need approval. Agents prepare freely; they ship only with your yes.

The autonomy ladder

Approval requirements are not fixed forever. Agents earn autonomy through clean runs: a track record of well-formed proposals and correctly executed approvals lets an agent graduate specific, low-consequence action classes to autonomous execution. Anything external and irreversible stays gated. Graduation is earned, never self-granted, and every action remains fully attributed in the audit trail regardless of how it was authorized.

What Havenware is not

It is not a chatbot with plugins, and it is not an autonomy demo. It is operational infrastructure: a team of specialists doing real work, under a control loop designed so that you can trust the output without doing the work yourself.

Continue with Agents & roles, or go straight to The approval queue for the governance contract in detail.