Trust & governance

Trust has a review path.

Every agent has a reviewer and an accountable person. Review paths, limits, and audit histories make that promise concrete.

There is setup. We walk you through turning on a department, bringing your brand, and placing the first gates so you are not dropped into a blank canvas. The destination is a team that is precise, systematic, and better over time, with you still holding the keys.

  • Ownership

    Your work belongs to you

    Everything your team produces belongs to you: drafts, documents, campaigns, records, and the System itself. CoSystem claims no ownership of your workspace content and no license to it beyond what is needed to run the service you asked for.

    That ownership carries the other way too. Because the work goes out under your company name, you are the party accountable for it, which is why the approval gates exist and why they are on by default.

  • Parity

    Permissions stay within your access

    An agent can never hold a permission you do not hold yourself. Scope is inherited from the human who owns the agent, and it narrows from there.

    Reducing your access also reduces the permissions available to your agents. An agent cannot retain access that the person who granted it no longer holds.

  • Gates

    Approval gates, with the action in front of you

    Actions that leave the workspace, sending mail, spending money, publishing, or writing to a customer system, are gated by default. A gated action stops before it happens and waits for you.

    You are not asked to approve a summary. You see the exact email, the exact charge, the exact record change, along with what the agent was trying to do. You approve the artifact you can see, and only that artifact ships.

  • Readback

    Work starts with a readback

    Before an agent starts consequential work, it restates the assignment as testable criteria. Where the work needs your approval, that readback is what you approve, the agent's understanding of the job, not your original phrasing.

    The finished work is then checked against the approved readback by a different agent. Done means the criteria were met, not that text was produced.

  • Stop

    A kill switch at three levels

    Stop the entire org, one agent, or a single recurring routine. Work in flight halts, queued work is dropped, and nothing resumes until you restart it deliberately.

    Stopping is never destructive. History, drafts, and the audit log survive, so you can read what happened before you decide whether to restart, re-scope, or remove the agent.

  • Log

    A full audit log

    Every action, approval, rejection, permission change, and configuration edit is recorded with its actor, timestamp, inputs, and result. Approvals record the human who granted them.

    The log is append-only. It cannot be edited or selectively deleted from inside the product, including by you, and it can be exported in full.

  • Export

    Your data leaves with you

    You can export at any time, and closing your account does not change that. The export covers workspace content, the System and agent configuration, approval history, and the audit log, in open formats.

    When you confirm deletion, removal propagates across our systems and to the subprocessors that hold copies, on the schedule in the privacy policy.

  • Secrets

    Encrypted tenant credentials

    The credentials you connect, mail, billing, storage, customer systems, are encrypted at rest with per-tenant keys and are never displayed back to you or written into logs, prompts, or exports in readable form.

    An agent receives the narrowest usable grant to a connected account rather than the credential itself, and revoking a connection cuts every agent off from it at once.

Accountability, in practice

Every decision has a record.

If a control on this page ever stops being true, that is a defect and we will treat it as one. Questions go to hello@cosystem.ai.