Contact

Program case study

Agent Platform

Governed AI agents with clear permissions and an auditable record of every action.

The operating layer being built to connect channels, agents, people, and systems, with scoped permissions, human approvals, and a record of every action — in development.

In development

  • Intelligent Systems
Visit agp-ui-pi.vercel.app All projects
Illustration of a request moving through a policy gate and human approval checkpoint to an audit ledger
Program artwork, not a screenshot of the software.

Research context

Agent Platform sits in the Intelligent Systems area. It asks how an agent can act across channels and systems while a person still holds the decision.

The program is in development, so this case study describes the operating layer it is building and does not present it as a deployed product.

Approach

  1. Give each agent only the permission a task needs, scoped and time-bound.
  2. Check the organisation's rules before an action runs, not after.
  3. Pause for a person when an action matters, and record what was decided.

System architecture

Diagram — how a request moves through the Agent Platform operating layer. Read top to bottom: a request arrives on a channel, the core applies routing, rules, and scoped permissions, a person decides when the rules require it, and the action reaches the target system with a record.

  1. Entry

    Channels

    The surfaces a request arrives on, from messaging to documents and the web.

  2. Core

    Routing, rules, and scoped permissions

    Connects channels, agents, people, and systems, and gives each agent only the permission the task needs.

    • Scoped permissions

  3. Decision

    Human approvals

    An action that matters pauses for a person, and the approval covers that action only.

    • Human approvals

  4. Record

    Action records

    Every action and outcome is recorded, so a reply can be traced to who asked and who allowed it.

    • Action records

  5. Target

    The systems agents act on

    The downstream systems an approved action reaches.

Current implementation

Published status
In development

The status above is the published one: the operating layer and its three named capabilities are what the program is building.

What this does not claim

  • Nothing on this page describes a production deployment, a customer, or a measured result.
  • The targets published on the program site are engineering targets, and the site says so itself; they are not results from customer deployments.

Evidence

What the record establishes

The evidence is the program's published description and its in-development status.

What it does not establish

There is no public deployment record yet, so no adoption or performance claim is made.

Roadmap

Every item below is planned, proposed, or under investigation. None of it is shipped, and this list is not a release schedule.

  1. Planned

    General availability

    Planned: the program site publishes engineering targets for general availability and states that they are targets, not customer results.

  2. Proposed

    A public evidence page

    Proposed: once the approval and record paths are open to external review, publish what they do and do not establish.

Collaboration

Agent Platform grows through teams that already run agents on customer-facing channels and want the rules and the record to be checkable.

Partner with us How the lab works

Work on a program with us.

Each program takes on partners where the research question meets real practice — clinical review for AIMed, agent runtime governance for ActantOS and Agent Platform, academic workflows for ScienzaOS, and document-heavy collections for AgenticOCR.

Discuss a collaboration
  • Clinical pilots
  • Research collaborations
  • Dataset partnerships
  • Student theses
  • Open-source contributions