Harness engineering

The model is the part of an agent that thinks. The harness is the part that lets it work.

The model is the brain. The harness makes it a teammate.

It is tempting to describe an agent as a model with a prompt attached. Real work quickly requires persistence, context, tools, and a place to keep results.

A useful agent needs to know what happened before, what matters now, which tools it can use, and when a consequential action should wait for a person. It needs somewhere to put files and results. It needs a way to continue a project after the first conversation ends.

The model still matters enormously. It interprets the request, reasons about the next step, and produces the language or media that moves the work forward. It is one component in a larger operating system for work. That surrounding system is the harness.

Five jobs a harness has to do

• 01 Persistence

  Keep sessions, files, decisions, and results available beyond one response.

• 02 Context

  Bring the right project knowledge into the run without replaying everything.

• 03 Skills

  Turn a broad goal into a repeatable method with instructions and examples.

• 04 Tools

  Give the agent useful capabilities while keeping permissions and approvals visible.

There is a fifth job running through all four: model selection. Different tasks call for different tradeoffs between capability, speed, cost, and modality. The harness makes that choice part of a flexible workflow.

Context needs a home

Chat history records what was said. Project context also holds the instructions, files, connections, and decisions that make the work coherent over days or weeks. A research agent should be able to return to its sources. A coding agent should be able to see the repository and the decisions around it. A writing agent should remember the audience, voice, and constraints without making you paste the same briefing into every new chat.

Dexto Projects give ongoing work a home, while Memory keeps useful preferences and decisions available across future work. Memory surfaces the context that matters when it becomes relevant.

Skills turn intent into a method

“Prepare a weekly update” names an outcome. A good result also requires a procedure: gather updates from several tools, check for missing information, apply a house style, and ask for approval before anything is sent.

Skills make that method reusable. They give an agent the instructions, examples, and supporting material needed to handle a specialized job consistently. They also let the harness load detail progressively, so every task starts with a clear operating context. This keeps the permanent prompt focused.

This is one of the quiet differences between a chatbot and an agent system. The system answers the current request and improves the path to the next one.

Tools need boundaries

An agent becomes useful when it can do things outside the chat window: read a project, update a tracker, send a message, create an artifact, or publish a site. Useful autonomy depends on clear boundaries around those capabilities.

Connections give Dexto access to the services people already use. Approvals keep consequential actions visible and let a person decide when the agent should continue. Read-only work can move quickly; actions that change data, publish externally, or manage access can pause at the right moment.

A strong harness makes these boundaries part of the experience. You can see what the agent is doing, understand why it needs a capability, and take over without losing the surrounding work.

The workspace is part of the product

Real work rarely fits inside a model response. It involves files, code, browsers, generated artifacts, and intermediate results. If those things disappear when the chat ends, the agent is always starting over.

Dexto gives agents a persistent cloud computer where they can store files, run code, open browsers, and use apps. That changes the unit of work from “answer this message” to “move this project forward.” It also means the agent can keep working when your laptop is closed or your machine lacks the hardware to run the task locally.

The workspace is where the agent's reasoning becomes something you can inspect, revise, and keep.

A good harness gets out of the way

The best harness makes a useful result feel straightforward while preserving the details that make the result trustworthy:

• the agent knows what it is working on;
• the relevant context arrives when it is needed;
• the model fits the task;
• tools are capable, with clear boundaries; and
• progress, files, and decisions remain available to the person doing the work.

That is the idea behind Dexto. The model supplies intelligence; the harness turns it into continuity.