Preface Executive Briefing

Week of May 23, 2026

From One Agent to an AI Team

Role design, governed memory, structured review: not one agent doing everything.

Reliable AI systems come from deliberate design: roles with clear boundaries, memory that is governed rather than unlimited, and agents organised into structured team workflows. This briefing explains the three engineering principles that turn a general assistant into a governed AI system.

Role Design

From one agent to roles

A role is not a persona. A role is a boundary: what the agent owns, what it must hand off, and what it is not allowed to decide.

The problem

One agent owns the whole chain

One entity · zero handoffs
  • Analyse evidence
  • Synthesise recommendation
  • Review risk
  • Approve
  • Execute

Both player and referee: nothing is held accountable to anything else.

The harness

The same work, split into five accountable roles

01
Analyst
cannot draw conclusions
02
Director
cannot approve or execute
03
Risk Reviewer
cannot approve
04
Approver
cannot execute
05
Executor
cannot change scope

Memory Governance

From memory to memory governance

Memory makes agents more useful, and more dangerous. What matters is not how much an agent remembers, but what it is allowed to remember, and in what form.

Raw Memory

"Customer was refunded immediately last month."

Memory Gate

Governed Memory

"One-time goodwill refund; not reusable as policy; approval required if repeated."

Five governance actions

Every memory becomes a policy choice

The gate does not just decide whether to remember. It decides the form that memory is allowed to take.

  1. 1

    Retain

    Keep exactly as is: accurate, current, safe to reuse.

  2. 2

    Compress

    Summarise the core fact; drop noise that does not change decisions.

  3. 3

    Expire

    Set a shelf life; refresh or discard after the window closes.

  4. 4

    Block

    Never let this enter working context; it must not shape behaviour.

  5. 5

    Escalate

    Flag for human review before the agent is allowed to act on it.

Operating Rhythm

From roles to AI teams

Planner

Tech Lead

Specialist

Coder

Reviewer

Code Reviewer

QA

QA Engineer

Human Lead

Engineering Director

The same role pattern shows up across every domain. The labels change; the structure does not.

Every AI team that works in production follows the same operating rhythm: plan, execute, review, approve. The roles below are not job titles. They are gates with explicit authority and an explicit handoff. The labels translate across industries; the underlying harness does not change.

In Practice

GStack: the AI-team pattern in practice

GStack is a real example of the pattern this briefing is teaching: a team of role-specialised agents, each with its own boundary. The transfer is the structure, not the tooling. The same role-gate pattern can govern banking research, customer operations, or any high-consequence workflow where AI moves from answer to action.

CEO

Finds the 10-star product hiding inside the request.

From Concept To Lab

n8n

See the harness in motion

If you would like to go deeper, we have prepared a pre-built n8n workflow so you can experience the pattern hands-on: role separation, evidence handoff, risk review, and decision rights operating as one team.

Book a Deep Dive

"A harness is not simply a better prompt. It is a designed system that splits AI work into roles, governs what memory can shape future behaviour, and organises agents into repeatable team workflows."

FAQ

More memory raises capability; it does not raise safety. Governed memory adds a gate that decides what each fact becomes: retained, compressed, expired, blocked, or escalated. The agent ends up remembering less, but remembering the right things in the right form.