You won't have 100 AI agents.

DeepSeek released something unusual this month.

Not another model.

A HARNESS.

(A prediction: this word will become as common as "agent" by next year.)

Near the top of the page is an equation:

Agent = Model + Harness.

Here is the simplest way to picture it:

The model is the engine. The harness is the racing car built around it. Together, they become the agent.

A stronger engine can help. But it still needs steering, brakes, sensors and a way to get power onto the track.

I used that same equation in Letter 79, almost in passing. It appeared near the bottom of a long letter about context engineering.

I still think context is a big story.

But I probably buried a bigger one in that letter.

For three years, most of us have treated the model as the product. We compared GPT with Claude, Claude with Gemini, one benchmark with another. Then we built a separate GPT, workflow, or agent for every job we wanted AI to do.

Sales agent. Research agent. Finance agent. Content agent. An agent that monitors the other agents.

That architecture is already getting old.

You probably won't have 100 AI agents.

You'll have a few trusted harnesses that know how to perform 100 jobs.

A harness, in normal language

A language model can produce a response.

It cannot open your files unless something gives it access. It cannot remember last month's work unless something stores that memory. It cannot use your CRM, ask for approval, recover after a mistake, or know when the job is finished on its own.

The harness does that.

In company terms:

The model is the intelligence.

The harness is the workplace built around it.

It gives the model a desk, tools, files, permissions, memory, rules, feedback and someone to call when things go wrong.

Claude Code and Codex are good examples. People call them coding agents, which is true. They are also sophisticated workplaces for models. They give the model access to files and tools. They preserve a working session. They show it the result of each action. They can ask for approval. They let it try, inspect what happened, fix the mistake and continue.

Change the model and you change the worker.

Change the harness and you change the job the worker is capable of doing.

That second change is where a lot of Applied AI is moving now.

How we got here

The first wave was custom GPTs.

Take a chatbot. Add instructions and a few documents. Give it a name. You now have a specialist sitting in a chat window.

Useful. Limited.

Then came AI workflows. The model became one step in a route chosen in advance. Read the form, classify it, draft the reply, send it for review.

The workflow knew the route. The model handled one part of it.

Agents changed that. Instead of following every step in a fixed order, the model could look at the situation and choose what to do next. Search first. Open a file. Call a tool. Ask another agent. Try again.

That freedom made agents more useful. It also made the surrounding system much more important.

If the agent can choose, somebody has to decide what choices exist.

What can it read? What can it change? Which tools are available? What survives into the next session? What requires approval? What counts as done? What happens after the third failed attempt?

Those are harness decisions.

The agent is the worker. The harness decides the conditions of work.

Then came skills

This is the part that changes the architecture.

A skill is a reusable set of instructions, resources and sometimes scripts that the harness can load when a particular job appears.

In normal company language, a skill is an SOP the AI can actually use.

Imagine one trusted harness connected to the systems your team already uses. It knows who you are. It knows what you are allowed to access. It keeps a record of what happened. It asks before taking a sensitive action.

Inside it, you might have a skill for preparing a customer briefing.

Another for reviewing a contract against your playbook.

Another for turning the monthly numbers into a finance memo.

Another for preparing an interview kit or a board update.

You do not need a new personality, memory system, permission layer and interface for each one. The workplace stays. The job manual changes.

OpenAI now publishes skill examples across marketing, sales, finance, legal, HR, operations and management. Anthropic supports the same basic idea across Claude.ai, Claude Code and its agent platform.

That matters because people still hear "harness" and think this is a coding story.

It isn't.

Coding is simply where the pattern became visible first.

Why code went first

Code already gives an agent a very good workplace.

The files are readable. The tools are available. Tests can say yes or no. Version control records every change. A broken edit can be reversed. An error message tells the agent what happened.

Most business work is nothing like that.

The real process lives across a policy document, three software tools, an old email thread and Nadia's memory of what happened the last time. Nobody has written down the exceptions. "Done" means somebody senior looked at it and felt comfortable.

There is no harness hiding in that mess.

This is why buying a stronger model often changes very little. The new worker arrives and the workplace is still broken.

The useful work is making the environment legible. Connecting the right systems. Writing down the procedure. Separating allowed actions from forbidden ones. Deciding where proof is possible and where a person still has to own the call.

That is not glamorous.

It is Applied AI.

What this changes for you

The next time somebody proposes a new AI agent, ask a different question:

Should this be an agent at all, or is it one more skill inside a system we already trust?

Take contract review.

The job-specific part is the review procedure: which clauses matter, which fallback positions are acceptable, which sources can be cited, and when the work must stop for a lawyer.

That is the skill.

The common part is identity, access, memory, logging, approvals and the connection to the document system.

That is the harness.

Mix those two layers together and you rebuild the same infrastructure for every use case. Keep them separate and every new skill becomes cheaper to add, easier to inspect and easier to move.

There is a limit, of course.

Some work should remain a fixed workflow. Some jobs need a dedicated application. A folder of instructions does not turn uncertain work into safe automation. Skills still need testing, permissions and human judgment.

But a large number of the "AI agents" companies are building today are really the same general agent wearing a different instruction file.

We do not need to keep rebuilding the worker.

We need to teach the workplace one more job.

One thing to do Monday

List the GPTs, automations and agents your team already uses.

For each one, split it into two columns.

In the first: everything that should be common. Identity, company context, tools, permissions, memory, approvals, logs.

In the second: everything specific to this job. Instructions, examples, templates, boundaries and the definition of done.

The first column is your harness.

The second is your skill.

If every use case has rebuilt both columns from zero, you do not have an AI system yet. You have a collection of experiments.

The part after the plus sign

DeepSeek made models swappable inside its harness. Tools are swappable. Skills, memory, sessions, sandboxes and even the agent loop are swappable.

That is a technical design choice, but it points at something larger.

Models will keep changing. The one you use next year may not exist today.

The durable layer is the part that knows your work. Your systems. Your permissions. Your procedures. Your standards. Your evidence. The points where a human must still decide.

I buried the equation once, so let me end with it this time.

Agent = Model + Harness.

Everyone will rent the model.

Most of your advantage will sit after the plus sign.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)