Harness Engineering

At a Glance

The Good Old Days

Old-school LLM chat was simple. You threw a prompt at the LLM, and the LLM returned an answer.

flowchart LR
  You[You] -->|Prompt| LLM[The LLM]
  LLM -->|Answer| You

  classDef human fill:#102033,stroke:#00f0ff,color:#e8f4ff,stroke-width:2px;
  classDef llm fill:#302500,stroke:#ffb000,color:#fff4d6,stroke-width:2px;
  class You human;
  class LLM llm;

In that world, context was almost entirely assembled by hand inside the prompt.

Current

Today, an agent / harness with project context and tools stands in front of the LLM.

---
config:
  flowchart:
    htmlLabels: true
  themeCSS: |
    .nodeLabel { text-align: center !important; }
    .nodeLabel .llm-row { display: flex !important; flex-direction: row !important; justify-content: center !important; align-items: center !important; gap: 6px !important; margin-top: 4px !important; }
    .nodeLabel .llm-row img { width: 22px !important; height: 22px !important; object-fit: contain !important; }
    .nodeLabel .llm-row img.ms-ico { width: 17px !important; height: 17px !important; }
---
flowchart LR
  Project[You + Project] -->|prompts / instructions / skills / MCP| Agent[The Agent<br/>aka Harness<br/><br/>Copilot Chat<br/>Copilot CLI<br/>Claude Code<br/>Codex]
  Agent -->|answer / PR / edit| Project
  Agent -->|context| LLM["<b>LLM</b><div class='llm-row'><img src='/theomonfort/llm-openai.png' alt=''/><img src='/theomonfort/llm-anthropic.svg' alt=''/><img src='/theomonfort/llm-gemini.png' alt=''/><img src='/theomonfort/llm-grok.png' alt=''/><img class='ms-ico' src='/theomonfort/llm-microsoft.svg' alt=''/></div>"]
  LLM -->|next step| Agent
  Agent -->|tool call| Tools[Tools<br/>read / edit / run]
  Tools -->|result| Agent

  classDef human fill:#102033,stroke:#00f0ff,color:#e8f4ff,stroke-width:2px;
  classDef llm fill:#302500,stroke:#ffb000,color:#fff4d6,stroke-width:2px;
  classDef agent fill:#132812,stroke:#9bbc0f,color:#f4ffd8,stroke-width:2px;
  classDef tools fill:#2a0f1a,stroke:#ff2e88,color:#ffd6e7,stroke-width:2px;
  class Project human;
  class LLM llm;
  class Agent agent;
  class Tools tools;

No magic. The agent is a layer that manages what to read, which tools to use, and how to return the result — instead of calling the LLM directly.

Under the Hood: Agent / Harness (Simplified)

• Execution Loop: The LLM decides the next move, executes a tool → returns the result to context, and repeats until done.
• Context Management: Organizes the system prompt, available tools, user task, and tool results, passing them as context with each LLM call.

# --- Setup ---
system_prompt = "You are a helpful coding assistant..."
available_tools = [search_web, read_file, edit_file, run_terminal]

# --- Agent Loop ---
user_task = input("How can I help you?")
context = [system_prompt, available_tools, user_task]

while True:
    next_step = await llm.determine_next_step(context)
    context.append(next_step)

    if next_step.intent == "done":
        return next_step.final_answer

    result = await execute_tool(next_step.tool, next_step.args)
    context.append(result)

What to Harness With?

There is no single technology tool that makes AI powerful. Separate what to always load from what to call only when needed.

The GitHub Docs names are Repository-wide custom instructions and Path-specific custom instructions. On the VS Code side, the latter is also called file-based instructions.

Ecosystem Comparison

The same “AI scaffold” concept exists across ecosystems, but file locations and names differ slightly.

Copilot’s strength is native support for the formats of major vendors. In the CLI, type /help to see available formats and commands.

Common Concepts

A good harness is not just a collection of tools — it defines how AI should proceed without getting lost.

Designing spec · phase · file ownership · role/objective/deliverable · context · approval upfront makes AI not just faster, but produces fewer reworks too.

Example: Ultralight

Ultralight is a multi-agent orchestration example by Burke Holland, Developer Advocate at Microsoft.
It creates a multi-phase execution plan, detects file overlaps, and acts as a harness that distributes work in parallel to Planner / Coder / Designer.

flowchart LR
  User[User prompt] --> O[Orchestrator<br/>Claude Sonnet 4.6<br/>multi-phase plan]

  O --> P[Planner<br/>Claude Opus 4.6<br/>research + docs]
  O --> C[Coder<br/>GPT-5.3-Codex<br/>scoped code changes]
  O --> D[Designer<br/>Claude Opus 4.6<br/>UI / UX owner]

  D -.-> S[Frontend Design Skill<br/>used by Designer<br/>brand / layout / CSS]
  C -.-> M[MCP Server<br/>used by Coder<br/>GitHub / Playwright<br/>docs]

  P --> O
  C --> O
  D --> O
  O --> R[Pull Request<br/>human review]

  classDef host fill:#102033,stroke:#00f0ff,color:#e8f4ff,stroke-width:2px;
  classDef agent fill:#132812,stroke:#9bbc0f,color:#f4ffd8,stroke-width:2px;
  classDef harness fill:#2a1020,stroke:#ff2e88,color:#ffe8f4,stroke-width:2px;
  classDef ship fill:#302500,stroke:#ffb000,color:#fff4d6,stroke-width:2px;
  class O host;
  class P,C,D agent;
  class S,M harness;
  class R ship;

🚀 I made a Codespace-ready template repo so you can try it in a few clicks: theomonfort/ultralight-template

Harness for coding this website

This playbook site is built with the same kind of harness. The Orchestrator (Theo) ties everything together with instruction files and prompts, then switches between CLI built-in agents, custom agents, skills, and MCP per phase.

🟦 Orchestrator / Custom agent · ⬜ CLI Built-in · 🟩 MCP · 🟥 Skill. Local = configured on my machine, Repo = checked into this repository.