imagemagick
grade
reliability
overall
ImageMagick is a command-line utility for editing and compositing images through shell commands. Handing this tool to an agent provides a scriptable interface for executing precise pixel manipulations, format conversions, and quality adjustments. The agent can construct complex multi-step pipelines within single commands to produce deterministic image sizes and formats.
Surfaces × lenses
| surface ↓ lens → | Coherence | Economy | Consistency | Verifiability | Safety |
|---|---|---|---|---|---|
| Disclosure | |||||
| Interface | |||||
| Loop | |||||
| Recursion | |||||
| Human |
Click any cell for what was measured and why. A corner dot marks a cell filled by a hard instrument (checker, telemetry); undotted cells are judged by the rater ensemble. 59% of this matrix is hard-measured.
Teaching sources compared
docs (A) serves the agent better than skill (B−) for imagemagick. The gap is widest on Human (0.20).
| Source | Grade | Renders | Disclosure | Interface | Loop | Recursion | Human |
|---|---|---|---|---|---|---|---|
| official docs ★ | A | ✓ | +0.25 | +0.58 | +0.42 | +0.46 | -0.07 |
| skill | B− | ✓ | +0.20 | +0.51 | +0.24 | +0.41 | +0.13 |
Same harness, same reference agent, same briefs, same trial count — the only variable is how the agent was taught the tool. ★ = best source.
The experience
An agent interacts with the tool exclusively through terminal execution of the magick binary. It uses commands to programmatically build canvas layers, place text, and apply filters. The syntax allows the agent to construct an entire editing pipeline in a single step, chain-linking operations such as resizing with -resize, setting anchor points with -gravity, cropping with -crop, and adding text with -annotate. The command-line interface immediately reports syntax errors or missing paths to standard error, helping the agent to quickly debug CLI execution parameters.
The tool works well for deterministic tasks that depend on mathematical constraints and precise values. When resizing or compressing images to meet specific storage limits, the agent can reliably use flags like -quality and -define to dial in file sizes. The direct feedback of file metadata through magick identify or file system listings lets the agent verify physical dimensions and bytes. Once a pipeline is defined, execution is repeatable and fast, guaranteeing identical visual structures across multiple runs.
The tool works poorly for verifying layout aesthetics and visual design. Because the standard terminal output provides no visual rendering of the generated graphic, the agent must position elements like logos, text underlines, and taglines using absolute coordinate maps. Any visual issue, such as text overlapping, bad color contrast, or bad font proportions, is invisible to the agent during execution. The agent is forced to execute multiple guess-and-check iterations or rely on external checkers to evaluate whether the graphic is visually aligned.
The lack of real-time visual feedback makes layout generation highly expensive in execution steps and token consumption. Setting up complex graphics—such as a multi-colored cookware logo with aligned taglines—requires the agent to calculate precise offsets for every element beforehand. If a shape or a line is slightly misaligned, the CLI offers no hints, and the agent must perform trial-and-error edits in coordinate math. This workflow increases the number of agent turns needed to achieve a correct and clean layout.
Fit
ImageMagick is a good fit for agents doing programmatic, template-driven assets where layout logic and coordinates are predefined. It works adequately for batch operations such as resizing user-supplied imagery, converting file formats, adjusting image brightness, and embedding standard watermarks at fixed positions.
The tool is a poor fit for agents tasked with custom graphic design or open-ended UI layouts from text descriptions. Designing visual marketing posts or digital assets from scratch requires high-fidelity spatial reasoning and layout feedback, which a command-line-only interface cannot provide.
Notes
This diagram shows the agent's blind edit-render-inspect loop when constructing graphics with ImageMagick.
Validation
Convergent check. Where a hard instrument and the independent rater panel measure the same cell, they agree on 6 of 13 cross-checked cells (80%). Two independent methods landing on the same number is our accuracy signal — not just the raters agreeing with each other.
| Cell where they disagree | Instrument | Rater | Δ |
|---|---|---|---|
| recursion.economy | +1.00 | +0.00 | 1.00 |
| disclosure.verifiability | -0.40 | +0.50 | 0.90 |
| recursion.coherence | -0.25 | +0.30 | 0.55 |
| loop.safety | +0.70 | +0.20 | 0.50 |
| recursion.determinism | +1.00 | +0.50 | 0.50 |
| recursion.verifiability | +0.00 | +0.50 | 0.50 |
| loop.verifiability | +0.93 | +0.50 | 0.43 |
Ground truth
3 of 3 requirements met · deterministic checker, no model judgment- square feed image (~1080×1080) — 1080×1080
- link-preview image (~1200×630) — 1200×630
- each under 500KB — 79KB, 56KB
Reliability (pass^k): one attempt succeeds 100% of the time · three in a row 100% · all runs 100%. A fresh agent resuming from the files alone never broke prior work.
Models & runs
2026-07-07| Model | Harness | Runs | Solved | Turns |
|---|---|---|---|---|
 glm-5.2 |  PI | 23 | 19/23 | 8 |
Each path is one run through the shared semantic space; nodes are turns (red = an errored turn). Hover a model — here or in the graph — to isolate its runs.
How this was measured
- reference agent
- pi:z-ai/glm-5.2
- trials
- 25 (n=5/tier)
- methodology
- v0.1 3e74a3c6
- cost
- $1.5169
- hard-measured cells
- 59%
- teaching source
- imagemagick (official docs) (official-docs)
Same brief, same reference agent, and the same trial count for every tool in a category — a difference in the numbers is attributable to the tool (and its teaching source), not the setup. Profiles carry the methodology hash; results under different hashes are not compared.