Your AI agent doesn't just need managing. It needs orchestrating.

You're reading The AI Directive - practical AI intelligence for
business leaders. Subscribe here if
someone forwarded this.

Last issue was about AI agents needing a manager, not just a
prompt.

That still stands.

But there is a second discipline that sits next to management, and it
is one that leaders will hear more often as AI agents move from
experiments into real operations.

Orchestration.

It is an ugly word, but a useful one.

Management is about boundaries: what the agent is allowed to do,
where it must stop, who owns the output, what needs approval.

Orchestration is about flow: what happens after work is delegated,
how progress is tracked, how multiple pieces fit together, how the
system recovers when something stalls, and how a person knows whether
the work is actually moving.

In human terms, the analogy is simple.

You do not just tell a junior colleague, "sort this out", and then
hope for the best.

You give them the brief. You agree what good looks like. You ask when
they will come back. If the work goes quiet, you chase. If they hit a
blocker, you help them escalate. If the work depends on three other
people, somebody coordinates the hand-offs.

That is not glamorous work.

It is the work that stops things drifting.

AI agents have their own version of the same problem.

You ask for work to be done and it appears to start well. Then a
browser session drops. A gateway restarts. A task runs longer than
expected. The conversation is compacted. A reminder fires but the script
fails. One agent finishes a useful sub-task, but nobody picks up the
result. Another agent is waiting for approval, but the approval is
buried in a chat thread.

From the outside, the experience is familiar:

"Did you get my message?"

"Where did that get to?"

"Didn't we agree this last week?"

"Why has nothing come back?"

The difference is that with AI agents this is not just a human
follow-up problem. It is a system design problem.

If the work matters, the agent needs more than a prompt. It needs a
way to hold state, track commitments, report progress, resume after
interruption, escalate blockers and show what has happened.

That is orchestration.

Orchestration keeps delegated AI work moving

Orchestrator

assigns, tracks, resumes

Worker agents

research, draft, check

Human approval

judgement and stop gates

State + progress + proof stop the work disappearing into the system.

For senior leaders, this matters because it changes the operating
model.

The first wave of AI adoption was about individual productivity: one
person using one assistant to write, summarise or analyse.

Agentic work is different. It starts to look like delegated work
across a small system. One agent might research. Another might draft.
Another might check sources. A human might approve. A scheduler might
trigger the next step. A tracker might hold the commitments.

At that point, the question is no longer only "is the model
good?"

It is "who or what is coordinating the work?"

There will be cases where a person does that orchestration. There
will also be cases where an AI agent is the orchestrator, assigning
tasks to other agents and bringing the work back together.

That sounds futuristic. It is not as far away as it sounds. In
practice, it is just the next layer of delegation.

But there is a catch.

Different AI capabilities are good at different jobs. A coding agent
may be excellent at fixing a script and running tests. A research agent
may be better at comparing external sources. A planning agent may be
better at turning a messy brief into a structured sequence of actions. A
lightweight model may be fine for a mechanical check. A stronger
reasoning model may be needed for a strategic review.

Orchestration is where those choices become business design, not
technical trivia.

The wrong lesson is:

The better lesson is:

That does not mean every business needs a complex agent platform
tomorrow morning. Most do not.

It does mean leaders should stop thinking of AI work as a single
prompt-and-answer exchange.

If the task has several steps, several sources, a time delay, an
approval point, a hand-off, or a consequence if it goes wrong, then
orchestration is already part of the work. The only question is whether
you have designed it, or whether it is happening accidentally in chat,
memory and goodwill.

Accidental orchestration does not scale.

It also does not build trust.

The organisations that make agents useful will not only ask, "what
can the agent do?"

They will ask:

Who is coordinating the work?

Where is the task list held?

What happens if the agent goes quiet?

How does the system recover after interruption?

Who gets notified when something is complete or blocked?

Which decisions require a person?

This is where AI starts sounding less like a tool and more like an
operating model.

That is the point.

Take one AI workflow you are excited about and ask five plain
questions:

If the answer to most of those is "the person remembers", you have
not got orchestration yet.

You have delegation with fingers crossed.

1. Management sets the boundaries. Orchestration
keeps the work moving.

2. Agentic work needs state: task lists, owners,
deadlines, approval gates, progress updates and recovery paths.

3. Orchestration may become a role in its own right:
sometimes human, sometimes AI, often a blend of both.

If you are reviewing an AI agent idea this week, do not start with
the demo.

Ask what happens after the demo:

Where does the work live?

Who chases it?

How does it recover?

That is where the operating model starts.

See you Tuesday.

@TobySmithFirth on
YouTube

These are tools I use personally. Affiliate links marked - I earn
a small commission if you sign up, at no extra cost to you.

Some links in this issue are affiliate links. I only recommend
tools I actually use.