The Harness · The Token Department

Open a part of the graphic to explore it.

Agent = Model + Harness

Agent

ModelHot-swappable

Harness

InterfacesWhere you direct the agent and review its work. Skills and workflowsThe instructions and routines the agent follows. Context managersThe notes, files and history it draws on. Connectors and toolsHow it reads, searches and acts.

ExecutionWhere its tools run and its working files live.

The model proposes. The harness acts.

The model receives context and proposes a response or a tool call. The harness assembles the instructions and evidence, checks the relevant permissions, runs the tool and sends its result back to the model. This cycle can repeat several times before you receive an answer.

The harness also manages the conversation history and the amount of context sent with each request. Hooks can run checks at defined points in the workflow. These choices affect what the agent can do and how it behaves.

Design the harness for any model

The model should be hot-swappable. Keep your instructions, skills, context and tool connections independent of a particular model or provider. A common model interface, with adapters where needed, lets the harness work with different models. Choose the model for the task without rebuilding your working method.

Your harness holds your working methods

I think of a configured harness as reusable intellectual property (IP). The software is widely available. The useful difference comes from the instructions, workflows, context, evaluations and permissions you build around your work. For an organisation, that can include its research methods and quality checks.

Four model categories to understand

Choose the category that fits the task. These categories can overlap. A frontier model can also be open-weight.

Frontier models offer leading general capabilities for work such as analysis, writing and coding.
Open-weight models make their weights available to download and run under their licence. You can run inference on your own hardware or with a host.
Small models need less computing power. They can suit narrower tasks and on-device use.
System One models specialise in structured decisions, such as a category or score. TypeSafe's System One models are an example.

Persistence belongs with the agent's context and execution. It describes memory or state that continues across sessions and can work with different models.

My setup uses the same five parts

I choose a model category for the work. My harness holds the software, instructions, notes and tools that support it. The graphic follows the same five parts as the explainer above. The model is replaceable.

My setup

Agent

Model typesHot-swappable

FrontierBroad capabilityOpen-weightAvailable weightsSmallLower computeSystem OneStructured decisions

Harness

Codex

Antigravity

Muse

Cursor

Hermes Agent

Grok Bot

Interfaces

Skills and workflows

Context managers

ObsidianNotes and source material

QMDSearch through MCP

Connectors and tools

Xtwitterapi.io API

GitHubgh CLI and MCP

Playwright MCPBrowser tools

FMP APIFinancial Modeling Prep

Fiscal.aiFinancial research

Connections use an application programming interface (API), command-line tools or MCP (Model Context Protocol).

Execution

My MacOn-device tools and files

My serverHermes scheduled work

Managed cloudGrok Bot computer

A product can fill several roles in this map. Cursor provides an interface and harness software. Hermes Agent runs tools, manages memory and supports scheduled work. Telegram can be its interface, with my server as its execution environment. Grok Bot supplies a managed agent with chat or voice and a persistent cloud computer. The underlying model remains a separate choice in the Model box.

Obsidian holds my notes and source material. QMD helps find the relevant passages. pstack, Matt Pocock's skills, Impeccable and my own skills give the agent methods for different tasks.

My X connector is twitterapi.io. Hermes uses it for the scheduled digest, with X MCP through xurl as a fallback. GitHub is available through gh CLI and MCP. These connections supply tools the harness can call.

Tool execution and model inference are separate choices. My tools can run on my Mac while a provider runs the model.

A morning X digest uses all five parts

Here is an illustrative version of my scheduled research workflow. Hermes gathers relevant posts from X, uses a selected model to assess them and saves a digest. The harness supplies the working method.

Interfaces
I set the brief and review the result through Telegram. The interface gives me a way to direct the work.
Skills and workflows
A digest skill sets the relevance rules, source checks and output format. These are reusable parts of my research method.
Context managers
A stored watchlist and previous research notes give the task context. Obsidian holds the material. QMD retrieves passages when the task needs them.
Connectors and tools
twitterapi.io retrieves the posts. The harness runs that connector and saves the digest with source links. X MCP through xurl is the fallback.
Execution
Hermes runs the scheduled job on my server. The selected model assesses the supplied material. I review the digest before using it in my work.

Explore interfaces or building skills and workflows to start shaping your own setup.

The graphic is adapted from my AI Tools & Skills page. The explanation also draws on Matt Pocock's definition of a harness.