AI tools have just evolved and you need to know how.

If you're not building harnesses for your agents, do it now.

Rules came first. Then skills, then plugins, and now harnesses. Actually, harnesses have been around since the beginning, veiled in mystery by the teams who built them until recently.

Here is how these tools work together today and how to build with them so you are ready for 2027.

Here is where we are today.

Here is the cheat sheet to understand how rules, skills, hooks, plugins, and harnesses work with agents today.

• Rules: Define facts, conventions, and boundaries. Loaded when a session starts or when matching files open.
• Skills: Step-by-step workflows to perform tasks. Loaded on demand by an agent.
• Hooks: Run deterministic scripts over the work produced. Triggered automatically at specific moments.
• Plugins: A bundle of rules, skills, MCP servers, and hooks that can be enabled or disabled.
• Harness: A virtual 'office' for an agent. Loaded each turn in a conversation. Gemini, Claude, and ChatGPT web apps are harnesses. So are Antigravity, Claude Code, and Codex.
• Model: An AI brain.
• Agent: A model becomes an agent when it's working inside of a harness.

Let's talk about the rise of agent harnesses.

We've got something new coming.

An agent is just a model that's plugged into a harness.

At a 5th-grade level, a raw AI model is just a smart brain in a jar. Whoever builds the best body around that brain the fastest wins.

Most builders use Antigravity, Claude Code, and Codex every day to build apps. Those programs are harnesses.

You drop a model into one of them and it takes on the form dictated for it inside the app.

Antigravity is a harness and so are Claude Code and Codex.

I want to stress this point.

The Antigravity, Claude Code, and Codex programs you type into are interactive harnesses. When you pick a model from the dropdown menu inside one of those programs, that pairing creates an agent.

When you talk to that agent, it relies on its model brain to think and its harness body to read files, run terminal commands, and get things done for you.

I suspect models have internal harnesses, too.

It may sound weird at first, but I am convinced that frontier models have an internal harness baked right into their weights.

Some models think deeply. Some do not.

The deep thinkers take longer to reach an answer. They recheck their logic three or four times and test different ways to retrieve data before they speak. Something inside the model is triggering internal skills, rules, and verification loops.

I could go even deeper down the rabbit hole.

Many advanced reasoning models act like hidden orchestrators with subagents baked inside them. That would explain why today's top reasoning models burn through thousands of hidden tokens in seconds.

That is a theory for another day, though.

Harnesses can exist inside harnesses.

Think of harnesses like Russian nesting dolls, where a larger wooden doll holds a smaller doll inside it. A better analogy might be an office building that exists in a city that exists on the Earth.

Most of us have been working with software creation harnesses like Antigravity (e.g. City). But now it's easier than ever to create business-specific harnesses (e.g. News office).

So what happens if you slip an Antigravity agent into a custom Edgy News Office harness? The agent becomes a full news reporter with a sharp persona, press-pass research tools, source memory, and a deterministic linter that tracks the stats a reporter cares about.

Each layer inherits the layers inside it.

Building an agent harness is easier than ever

Antigravity agents know how to build them now. Here's what a finished one looks like in the filesystem:

newsroom-harness/
├── AGENTS.md                   # THE ENTRY POINT: Loads the newsroom harness on every turn
├── grades/                     # SUPERPOWER: Grading (agent must get passing grade to continue)
│   ├── check_permissions.py       # Blocks any tool call that isn't on the permission slip
│   ├── check_truth.py             # Checks facts and accuracy before finishing
│   ├── check_safety.py            # Scans files for security risks
│   └── check_reading.py           # Scores reading level and forces fixes
├── identity/                   # SUPERPOWER: Identity of the inhabitants who occupy the newsroom
│   ├── persona.md                 # Personality traits (curious, bold, careful)
│   ├── job.md                     # The exact work the agent focuses on
│   └── style.md                   # Voice, tone, and how it talks
├── memory/                     # SUPERPOWER: Long-term newsroom memory across sessions 
│   ├── memory_graph.json          # Important facts saved across sessions
│   └── dream.py                   # Background routine that cleans up old memories
├── permissions.md              # PERMISSIONS: Code of conduct when working in the newsroom
├── skills/                     # SUPERPOWER: Local skills specifically for the agent using this harness
│   ├── research-topic/            # How agent should research topics 
│   └── draft-script/              # Steps to create a draft while preserving previous drafts
└── tools/                      # SUPERPOWER: Custom scripts built just for the newsroom
     ├── video_producer.py          # A list of video production tools available to agent
     ├── generate_voiceover.py      # Calls an outside service only this agent uses
     └── save_to_gdoc.sh            # Calls a global 'gdocs' tool, but locks it to ONE Doc ID

The harness above represents a newsroom. When you attach a model to it, an agent is created that has access to this virtual newsroom. The agent becomes a newsroom agent and adopts the persona of someone who might work there (bold, curious, careful).

That's a lot of files in there! Not too long ago, this would have been a pain to create, but a modern agent can build this in 5-10 minutes. For a finished example, check out devrel_demos/agents/newsroom_harness on GitHub.

Now that you know what a finished agent harness looks like, let's actually build one and attach a model to it. This will produce an agent with unique superpowers!

Build an agent with unique superpowers.

Until late 2026, building an agent just meant writing a short note and handing the AI a few tools.

Today, you can build agents that are more capable than the ones we've seen in 2026. To do so, you need to first build an agent harness. The great news is that you can build them right inside programs like Antigravity and Codex. Quickly.

1. Determine what kind of harness you want to create.

What kind of office do you want to provide agents who embody your harness?

• A newsroom?
• A physics lab?
• A classroom?
• A landscaping business?

2. Determine what kind of agent superhero you want.

Who do you want an agent to become when it slips on your harness?

• What is its Big 5 personality? 
• What are its habits?
• What is its communication style?
• How often does it ask for human oversight?

3. Choose the agent's superpowers.

Choose the superpowers you want your agent to have while it's in the harness. Here are a few examples:

• Do you want to store memories in the harness?
• Do you want the harness to clean up memories on a timely basis?
• Do you want accuracy, reading, and security grading systems to ensure the quality of work produced?
• Do you want the agent to have special skills and tools that are not available outside of the harness?

4. Build the harness.

Now that you've got the general details collected for your harness, you're ready to build it. Load up Antigravity or a similar agentic program, select your most advanced model (Gemini 4, Claude 5.5, GPT-6), and give it a prompt that looks like this:

Hey! I need you to build an agent harness for me. It needs to be saved in <your_directory_of_choice>. It needs to include <list_your_harness_details_here>.

And that's it! The agent will build the harness.

5. Turn the agent who just built the harness into a superhero agent.

Once the harness is created, ask the agent who created it to put it on and then give it a task to perform. Maybe you want it to investigate a newsbeat, perform some physics research, organize 100 student papers, or create a landscaping presentation and 3D walkthrough for a client.

That's enough for today, have a good day.

For the last few years, we treated AI like a smart stranger we had to re-train from scratch every morning. We wrote longer prompts, crossed our fingers, and hoped the model would not forget our rules halfway through the chat.

Those days are over.

You do not need to write complex code to build a great agent anymore. You just need to be a good architect. Pick the office you want to create, choose the personality and superpowers that belong inside it, and ask your agent to build the harness for you.

Ten minutes later, that same agent can step right into the suit it just built and get to work.

Go build your first newsroom, physics lab, or dream workshop today and see what happens when a smart brain finally gets a real home!

Etc

If you want to try out the harness I created for this article, see devrel_demos/agents/newsroom_harness.

If you want to learn more about Antigravity, download it here: https://antigravity.google/.