Executive briefing

Week of 15 August 2026

From Agent Portfolios to Connected Workflows

Graph Engineering, and why the next stage of enterprise AI is coordinating the agents you already have

AI adoption is expanding faster than workflow ownership

The first wave of enterprise AI was built by individuals. Employees inside Microsoft Copilot, Copilot Studio and Google Gemini built assistants to search documents, summarise information, draft content and review data, and some teams went on to run several of them for one area of work. One person can coordinate several agents through tacit knowledge. Once the workflow crosses operations, technology, risk, compliance, finance and client teams, the organisation needs explicit ownership, handoffs and control points.

One person, three or four agents

They know what each agent is for, they can tell when an answer is wrong, and they fix it themselves. Coordination is personal, and it works.

Six departments, one workflow

Each department has its own agents, approvals and idea of what finished means. The organisation needs a way to represent the workflow, connect the capabilities in it, and assign responsibility for the outcome.

A signal from the agent-building community

Two words are doing all the work in that conversation, so here they are, drawn. Once you can see the difference between them, the rest of this briefing is about one of them.

One agent, going round

The agenttrychecktry again

An agent is given a task and works at it in a cycle: it tries something, checks its own result, and tries again until it decides the task is done. Everything happens inside one box, and one person watches that box. Almost every AI assistant in your organisation today works this way.

Many steps, joined up

when something is wrong, go backor escalate to a specialistStarta stepAgenta stepPersonapprovesDonehandoff

Draw several of those boxes and the connections between them and you have a graph. Each box is a step, whether an agent, a person or a system does it. Each arrow is a handoff. Some steps are approvals, and some arrows go backwards, returning the work to an earlier step or escalating it to a specialist.

Graph Engineering is the practice of drawing that second picture for a piece of work that matters, and then governing it: deciding which steps exist, who owns each one, where approval is required, and what happens when the work has to go backwards.

Posted on X
Are we still talking loops or did we shift to graphs yet?
Peter Steinberger@steipete17 July 2026

Creator of OpenClaw, now working for OpenAI

The signal is that the people building these systems have stopped talking about the first picture and started talking about the second one. That is all it is. It does not show that graphs are universally better, or that the term has settled into a standard. What it marks is a change of subject, and the same change is now arriving inside Microsoft and Google's enterprise products.

Most organisations are somewhere between an inventory and a portfolio

The third stage is the one that carries a business outcome, and it is the one almost nobody has reached.

Agent inventory

You know which agents exist and where they are being used.

Agent portfolio

Development is standardised, permissions are managed, acceptable use is defined, and someone decides which agents deserve further investment.

Connected workflow

Agents, systems, employees, approvals, evidence and exception paths work together across boundaries, under one business outcome.

Where the difficulty starts

Portfolio language can hide this: launched use cases, trained employees and a governance committee are real signals of investment, and none of them is evidence that work crosses the organisation reliably. The distance that matters is between knowing where your agents are and connecting them into work the organisation can govern.

Graph Engineering makes hidden workflow dependencies visible

A single reporting workflow, drawn as the business experiences it. Data arrives from internal systems, two agents do real work inside it, and people decide, review and approve around them.

Agent only
Agent with a person
Person only
Connected reporting workflowFive workflow steps run top to bottom through agents and people. Approval gates the final send, escalation reaches a specialist, and a failed check returns the work to an earlier agent step.or escalate to a specialistif the check fails, back to the agentSTEP 1Pull the dataSTEP 2Draft the readSTEP 3DecideSTEP 4Build itSTEP 5Check and send
  1. Step 1

    Pull and clean the data

    Figures are extracted from the finance system and prepared for use.

    Who IT and operations

    Runs on Client database

  2. Step 2

    Draft the first read

    An agent works through the data and writes the quantitative commentary.

    Who Analyst with the agent

    Runs on Copilot agent, reporting

  3. Step 3

    Decide what to act on

    A person reviews the insights and confirms which actions are required.

    Who Analyst

    Runs on Human judgement

  4. Step 4

    Build the dashboard

    A second agent turns the confirmed insights into a specification and a working preview.

    Who Analyst with the agent

    Runs on Copilot agent, build

  5. Step 5

    Check and circulate

    Dashboard quality and data consistency are reviewed before anything is sent.

    Who Leadership review and compliance

    Runs on Email and Teams

The organisation describes these as separate use cases. The business experiences one workflow.

Every arrow between two adjacent steps is a handoff, and each one raises the same four questions: does the complete process have a clear owner, are the handoffs controlled, is there visible evidence of what happened, and is there a reliable path when an exception appears. Those questions are unanswerable while the steps are managed as separate projects.

Illustrative. An anonymised design from a private-bank discovery conversation; deployment is unconfirmed.

Microsoft and Google are making agents easier to connect

Both vendors are shipping the connective layer, and it is worth knowing what has already arrived in products your organisation may be licensing today.

Microsoft

Copilot Studio agent flows

Flows that can be started by an event, a schedule, or another agent, and that run across your connected systems.

  • Connectors
  • Branching
  • Loops
  • Human approvals
  • Child flows
  • Monitoring
Microsoft Learn, agent flows overview
Google

Gemini Enterprise

The same direction from the other side: agents that delegate to other agents, under a shared identity and governance layer.

  • Graph-based sub-agent orchestration
  • Agent-to-agent delegation
  • Identity
  • Governance
  • Observability
Google Cloud blog, Gemini Enterprise

A platform can connect two agents. It cannot tell you who owns the result.

Ownership, data access, exception approval and stopping conditions stay organisational decisions. This is why the useful skill is identifying which connections a consequential workflow actually depends on, and governing those, rather than connecting everything that can be connected.

Leaders should build a portfolio of owned, bounded workflows

An agent owner is responsible for one agent: its purpose, tools, boundaries and evaluation. Several agents contributing to one outcome need something above that.

  • A workflow owner

    Accountable for the end-to-end business result, across every department the work passes through.

  • A control owner

    Defines the permissions, approvals, evidence, escalation and rollback the workflow has to satisfy.

  • A platform owner

    Provides identity, reusable interfaces, monitoring, versioning and lifecycle standards, so each team is not inventing its own.

Start with one consequential workflow and five questions

Deliberately small. Choose one workflow that matters, such as the reporting case above, and answer these before connecting anything.

01

What outcome is the workflow responsible for?

Name the business result, not the tooling. If it cannot be stated in a sentence, the workflow is drawn too wide.

02

Which agents, systems and people participate in it?

The full list, including the steps that are entirely human and the systems nobody thinks of as AI.

03

Where do approvals, exceptions and human escalation occur?

The control points. Mark where the work is allowed to stop, and who is allowed to stop it.

04

Who owns the end-to-end result?

One name. Owners of individual steps do not add up to an owner of the outcome.

05

What evidence would justify expanding the workflow?

Decide this before the pilot runs, so the answer is not written backwards from whatever the pilot produced.

Workflow improvement requires measures of changed work

A pilot should be judged on whether the work itself changed. The easy numbers describe how much AI is present, which is a different question.

Measures of changed work

  • Cycle time. How long the workflow takes end to end, not how fast one step got.

  • Review burden and rework. How much checking and redoing the work still needs.

  • Exception visibility. Whether you can see the cases that went off the path, and what happened next.

  • Explain and reverse. Whether an action can be accounted for afterwards, and undone.

Measures of activity

  • Agent counts. How many exist.

  • Licence adoption. How many people have access.

  • Usage volume. How often something was run.

What leaders should take away

The next phase of enterprise AI will be decided by whether organisations can connect the right capabilities around the right workflows while keeping responsibility visible.

01

The constraint has moved

It is becoming less about whether you can create another agent, and more about whether you can coordinate the ones you already have.

02

Draw one workflow before connecting anything

Where the work begins, which agents and people contribute, what must be checked before the next step, who can approve an action, and what happens when the result is wrong.

03

Cross-department work needs an owner above the agents

A workflow owner for the outcome, a control owner for permissions and escalation, a platform owner for the standards everyone builds on.

04

Fewer connections, better governed

Each connection you add has to be permissioned and maintained, and widens the set of steps a failure can reach. Connect what a consequential workflow depends on, and leave the rest unconnected.

05

Judge the pilot on changed work

Cycle time, review burden, exception visibility and reversibility. Agent counts and licence adoption measure activity.

FAQ

A way of describing a workflow as the connected thing it actually is: the steps, the handoffs between them, the points where approval or escalation is required, and the paths that loop back when something is wrong. The vocabulary comes from the people building agent systems, but each term maps onto something a leader already manages.