Harness engineering

Is GitHub Copilot a harness?

Two products under one name, and only one of them is a harness.

Download for macOS
v0.1.33 · .dmg · Apple Silicon & Intel
Signed & notarized by Apple · opens without a Gatekeeper warning
sha256 698955a0187bc039f4c74f5d05a9f10fbb27376a45788a0a241d1326b73873c7
Download for Windows instead
$curl -fsSL https://lucidtrain.com/install.sh | sh

It depends which Copilot. Inline completion is not a harness: it is a single-shot prediction with a short context window and no tools, no loop and no verification. Copilot's agent mode is a harness, because it assembles context, calls tools, iterates, and asks for permission. The product name covers both, which is why the question does not have one answer.

Completion is not a harness

Inline completion takes the surrounding code, predicts a continuation, and stops. There is no loop, no tool call, no permission decision and nothing checking the result other than you. It is a very good autocomplete, and none of the properties that define a harness are present.

Agent mode is

Agent mode assembles context across files, calls tools, iterates toward a goal, and requires approval for some actions. The VS Code team's own definition of a harness, the layer that assembles context, exposes tools, runs the agent loop, interprets tool calls and turns model output into something useful in the editor, describes agent mode precisely.

The useful test

Ask whether the thing can take more than one action toward a goal and change course based on what happened. Completion cannot; it emits once. An agent mode can, and the machinery that lets it is the harness. That test is more reliable than arguing about product names.

Where a harness sits relative to Copilot

They occupy different layers and are commonly used together. Lucid Train works above the editor: it generates the architecture diagram from your repository, uses it as an implementation specification, routes different roles to different models, and runs sub-agents in isolated git worktrees. It does not do inline completion and does not try to.

Which Copilot is a harness

Inline completionAgent mode
Multi-step loopNoYes
Tool callsNoYes
Context assembly across filesLimitedYes
Approval policyNoYes
Verification of its own outputNoPartial
Is it a harness?NoYes
FAQ

Questions

Is GitHub Copilot a coding harness?

Agent mode is. Inline completion is not, because it has no loop, no tools and no verification.

What makes something a harness rather than an assistant?

The ability to take more than one action toward a goal and adjust based on results, plus the machinery around that: context assembly, tool definitions, approvals and verification.

Can I use Copilot alongside a harness like Lucid Train?

Yes, they do not overlap. Copilot completes the line you are writing; Lucid Train works at the architecture layer and hands a diagram to a coding agent as a specification.

Related