# Agent Harness Landscape Visual Prompts

These prompts generated the three raster figures used by the Agent Harness Map. Each image was generated as a separate asset, reviewed visually, resized to 1600 x 900, and converted to WebP.

Selected Gemini source files:

| Published asset | Selected source |
|---|---|
| `agent-harness-160-not-runtimes.webp` | `Gemini_Generated_Image_21edgl21edgl21ed.jpeg` |
| `agent-harness-four-layers.webp` | `Gemini_Generated_Image_u0ieeju0ieeju0ie.jpeg` |
| `agent-harness-selection-funnel.webp` | `Gemini_Generated_Image_7rttz17rttz17rtt.jpeg`, with the pilot count corrected from 5-to-10 to 8-to-12 real tickets before conversion |

## Shared art direction

Use a warm cream graph-paper background (`#f5f1e8`), a subtle grid (`#d4cfb8`), precise dark pencil or ink lines (`#1a1a1a`), and restrained yellow (`#fbbf24`), green (`#16a34a`), and orange (`#d97706`) accents. Reserve red for risk gates. Keep the layout editorial and technical. Do not use photos, 3D, mascots, decorative doodles, stars, page numbers, or logos. All visible text is English. Target 1600 x 900 pixels.

## 1. Broad catalog versus strict runtime map

Output: `guide/images/agent-harness-160-not-runtimes.webp`

Alt text: A three-stage funnel separates a broad catalog of 160 projects in 12 categories from a loop-ownership test and a smaller strict runtime map, with adjacent and unknown projects kept apart.

```text
Create a clean editorial infographic for a technical documentation guide, landscape 16:9 at 1600x900. Visual language: warm cream graph-paper background (#f5f1e8), very subtle grid (#d4cfb8), precise dark pencil/ink lines (#1a1a1a), restrained yellow (#fbbf24), green (#16a34a), and orange (#d97706) accents. No photos, no 3D, no mascots, no decorative doodles, no stars, no page numbers, no logos. High information density but ample whitespace. All visible text must be exactly in English and spelled correctly.

Headline, large: “160 PROJECTS ≠ 160 RUNTIMES”
Under it, create a left-to-right evidence funnel made of three clearly separated cards connected by thin arrows:
1. Card title “Broad catalog” with large number “160”, small line “projects”, and a compact badge “12 categories”. Show a mixed stack of small abstract tool cards to signal heterogeneous entries.
2. Card title “Loop ownership test” with a simple circular agent loop diagram labeled “plan”, “act”, “observe”, “repeat”. Add a filter gate with the question “Who owns the loop?”
3. Card title “Strict runtime map” with a compact set of selected runtime blocks, plus two side bins labeled “Adjacent control planes” and “Unknown”.
At the bottom, add one concise note: “Catalog size is not runtime count.”
Make hierarchy crystal clear. Avoid tiny text. The image must remain readable embedded in a documentation page.
```

## 2. Four layers and responsibilities

Output: `guide/images/agent-harness-four-layers.webp`

Alt text: A four-layer stack separates the model, repository harness, runtime harness, and orchestrator, with control flowing down and evidence flowing up.

```text
Create a clean editorial infographic for a technical documentation guide, landscape 16:9 at 1600x900. Visual language: warm cream graph-paper background (#f5f1e8), very subtle grid (#d4cfb8), precise dark pencil/ink lines (#1a1a1a), restrained yellow (#fbbf24), green (#16a34a), and orange (#d97706) accents. No photos, no 3D, no mascots, no decorative doodles, no stars, no page numbers, no logos. High information density but ample whitespace. All visible text must be exactly in English and spelled correctly.

Headline: “FOUR LAYERS, FOUR JOBS”
Build one clear vertical architecture stack of four wide horizontal layers. Use thin arrows to show control flowing downward and evidence flowing upward.
Layer 4, top, orange accent: “ORCHESTRATOR” and subtitle “coordinates agents or runtimes”. Include three small connected worker nodes.
Layer 3, green accent: “RUNTIME HARNESS” and subtitle “owns plan → act → observe → repeat”. Include a concise circular loop with labels “plan”, “act”, “observe”, “repeat”.
Layer 2, yellow accent: “REPOSITORY HARNESS” and subtitle “instructions, hooks, skills, policy”. Include compact file tabs labeled “AGENTS.md”, “hooks”, “skills”.
Layer 1, bottom, neutral accent: “MODEL” and subtitle “generates and reasons”. Include one abstract neural block, not a brain illustration.
At the right, add a slim responsibility panel titled “ASK FIRST” with four questions aligned to the layers: “Who coordinates?”, “Who owns the loop?”, “What configures the repo?”, “Which model runs?”
At the bottom, add: “Do not compare tools from different layers as substitutes.”
Make the layer boundaries and ownership distinction unmistakable. Avoid tiny text.
```

## 3. Selection and evidence funnel

Output: `guide/images/agent-harness-selection-funnel.webp`

Alt text: A five-step selection flow defines the job, checks loop ownership, verifies evidence, pilots 8 to 12 real tasks, and ends in adopt, adjacent layer, or reject.

```text
Create a clean editorial decision infographic for a technical documentation guide, landscape 16:9 at 1600x900. Visual language: warm cream graph-paper background (#f5f1e8), very subtle grid (#d4cfb8), precise dark pencil/ink lines (#1a1a1a), restrained yellow (#fbbf24), green (#16a34a), and orange (#d97706) accents. Use red only for a risk gate. No photos, no 3D, no mascots, no decorative doodles, no stars, no page numbers, no logos. All visible text must be exactly in English and spelled correctly.

Headline: “CHOOSE THE LOWEST LAYER THAT SOLVES THE PROBLEM”
Create a left-to-right selection funnel with five large steps connected by arrows:
1. “DEFINE THE JOB” with four selectable labels: “Generate”, “Run”, “Configure”, “Coordinate”.
2. “CHECK LOOP OWNERSHIP” with the question “Does it own plan → act → observe → repeat?” and two paths “Yes” and “No / unknown”.
3. “VERIFY EVIDENCE” with a small checklist: “Official URL”, “README or docs”, “Security boundary”, “Recovery behavior”.
4. “PILOT” with a bold band “8–12 REAL TASKS” and three measured outputs: “success”, “cost”, “recovery”.
5. “DECIDE” with three outcome cards: “Adopt”, “Keep as adjacent layer”, “Reject”.
Below the funnel, add a horizontal red risk gate labeled “Human approval for consequential actions”.
At the bottom, add a concise note: “Features and popularity do not replace task-level evidence.”
Keep each step visually distinct, readable, and practical. Avoid tiny text.
```

## Verification record

| Asset | Dimensions | Size |
|---|---:|---:|
| `agent-harness-160-not-runtimes.webp` | 1600 x 900 | Selected Gemini output, resized and converted to WebP |
| `agent-harness-four-layers.webp` | 1600 x 900 | Selected Gemini output, resized and converted to WebP |
| `agent-harness-selection-funnel.webp` | 1600 x 900 | Selected Gemini output, corrected to the 8-to-12-ticket protocol, resized, and converted to WebP |