Andrew Osula AO

Research case R1

AI · Orchestration · Governance

Research

Governed AI Agents & Agentic Workflows

Exploring how AI agents can work across tools and structured context without removing human accountability.

AECO research / governed agents

Useful assistance starts with a defined task and ends with an accountable decision.

Conceptual architecture for design and delivery teams. Agent roles describe research directions, not deployed autonomous capabilities.

Project intelligenceBrief → coordinated review → controlled output
Research architecture
Task
orchestratorScope · route · record
03 / Human approval boundaryA person reviews the evidence and authorises the next action.
Draft until approved
01Define task02Select context03Use scoped tools04Validate evidence05Human review06Retain audit trail

The challenge

AI capability becomes difficult to govern when context, tool access, evidence and approval boundaries are implicit.

Teams are beginning to use AI across documents, applications and delivery systems. The difficulty is keeping the source of information, tool access and approval route clear when a task passes between them.

The approach

I am testing a task-led pattern: define the request, supply bounded context, restrict tool access, retain evidence and put a person at the decision point. The work compares how this pattern can be applied across Codex, Claude and Copilot environments.

System

Select a step to see its role in the work.

  1. 01Task

    Define a bounded task and the decision a person needs to make.

  2. 02Context

    Provide the approved sources and constraints needed for that task.

  3. 03Model

    Select a model for the task while keeping its output open to review.

  4. 04Tools

    Limit each tool to the access needed and record what it does.

  5. 05Evidence

    Retain source references and outputs so a reviewer can trace the result.

  6. 06Human approval

    Require a named person to approve consequential actions or publication.

  7. 07Audit trail

    Keep a record of inputs, actions and approvals for later review.

Current position

This is research into a reusable operating method. It is not presented as a live organisational deployment or a measured productivity result.

Research and prototyping, not represented as a production deployment.

Next question

The next test is whether the same controls remain understandable when a workflow crosses several tools and a reviewer needs to reconstruct what happened.

Knowledge graph

Follow the methods and information shared by this work and the cases below.

CurrentGoverned AI Agents & Agentic Workflows
Applied agent workflowACC Review AgentResearch · ActiveStructured AI interactionDesign Prompt StudioResearch · ResearchAI workflow continuityAI Chat Thread ManagerResearch · ResearchDecision-support opportunityResource Planning & Executive Decision SupportWork · Active