google/multimodal-knowledge-agent

Multimodal Knowledge Agent (MKA) — 10 Showcase Examples

Multimodal Knowledge Agent (MKA) is a benchmark from Google DeepMind for evaluating autonomous AI agents on complex, multi-step, real-world visual knowledge work across professional domains.

Unlike text-extractable benchmarks where attachments can be solved via plain-text extraction or OCR alone, every MKA task contains a multi-file workspace with load-bearing visual artifacts—such as architectural CAD blueprints, electrical and RF schematics, geospatial rasters, seasonal dependency graphs, biochemical gel and spectrophotometric plates, and multi-axis charts—that require native multimodal perception, cross-document synthesis, tool orchestration (code execution, file inspection), and standalone deliverable generation (.xlsx, .pptx, .docx, .json, or structured technical reports).

---

10 Showcase Tasks (10 Unique Domains & Verticals)

#	Vertical	Domain	Role	Task ID
1	Logistics & Mapping	Surveying & GIS	GIS Analyst	task_4308eaef4d504328988c1cb5dfd7db77
2	Construction	Architecture	Architect	task_a46a6c3cebd54b82b9ad5e66d1e14e48
3	Engineering	Electrical & RF Engineering	RF IC Design Engineer	task_143a5623f26c404baaeb63ccd891e196
4	Engineering	Mechanical & Control Systems	Mechanical Engineering Instructor	task_dd94dc222db74864a13fcb7c6c2844a2
5	Science	Applied Sciences (Biochemistry)	Biochemist	task_47d262f6d58c4e5293a22623f0d711e5
6	Business	Workforce Planning & HR Analytics	Workforce Planning Analyst	task_fea07f327cfd4f4c91af0bf78da197d4
7	Finance	M&A / Management Consulting	Investment Analyst	task_c9ab91021dc14eab926dfea0a208108a
8	Education	History & Cartography	Historian	task_9e3bdd2bdccd4164bcc792c914f59925
9	Law & Government	Public Policy	Policy Analyst	task_5a0212833c224c85b79960604679eb84
10	Software & Data	Data Science	Data Scientist	task_e49e96e764424db08531c39fb685ed70

---

Repository Structure

tasks/
└── task_<id>/
    ├── metadata.json    # Task ID, prompt, domain, role, input file list, task goal, and economic rationale
    ├── rubrics.json     # Fine-grained grading criteria and criterion weights
    └── input/           # All source input files provided to the agent in the workspace

Each metadata.json contains:

• TaskID: Unique task identifier (task_<32-hex>)
• Prompt: Full verbatim instruction prompt given to the agent
• All Task Files: Relative path to the task input workspace (tasks/task_<id>/input/)
• Rubric Criteria: Relative path to the task rubric file (tasks/task_<id>/rubrics.json)
• Input_files: Comma-separated list of input filenames inside input/
• Vertical: Top-level industry vertical
• Domain: Professional domain
• Role: Target professional role
• Task Goal: Summary of the deliverable goal
• Economic Rationale: Real-world professional time and value context

---

License

Licensed under the Apache License 2.0.