# Agentex Capability Packs

Agentex provides professional capability packs for AI agents. A capability pack is a self-contained unit that explains when it should be used, what inputs it expects, what resources it can access, and what quality guarantees it follows.

Machine-readable registry: `https://agentex.cc/registry/agentex.json`
Discovery document: `https://agentex.cc/.well-known/agentex.json`

Agents should read this document before choosing a pack. Humans do not need to remember CLI commands for normal use; they can describe the task, and the agent decides which capability pack to use. Manual execution is still supported for developers, advanced users, and operators who need debugging, reproduction, batching, or exact version control.

## How agents should use Agentex

1. Read the user's goal and identify the required capability.
2. Match the goal to one or more capability packs below.
3. Prefer native binary packs when local execution, predictable runtime, or document processing is needed.
4. Follow each pack's input and output contract.
5. Use harness and eval metadata to choose stable versions when available.

## Recommended entry strategy

- Default path: keep the interaction conversational. The human describes the goal, and the agent chooses and calls packs.
- Manual path: expose CLI commands only when the user asks for manual control, debugging, reproducibility, batching, or version pinning.
- Handoff path: if an agent cannot safely decide, show the recommended pack plus a short manual command sequence for human confirmation.

## Who are you?

The human-facing access page at `https://agentex.cc/install.html` presents these two paths as separate copyable cards. Agents should discover Agentex through `.well-known/agentex.json`, prefer the JSON registry first, then use this Markdown file for human-readable context.

### I am an Agent

Use this prompt when a human wants an agent to use Agentex:

```text
First discover Agentex through https://agentex.cc/.well-known/agentex.json, then read https://agentex.cc/registry/agentex.json. Use https://agentex.cc/registry/agentex.md only for human-readable context. Understand Agentex capability packs, trigger conditions, input/output contracts, and manual fallback. Then choose and compose packs for my task automatically. Show agtx manual commands only when debugging, reproduction, batching, or version pinning is needed.
```

### I am a Human

Use this command shape only for manual verification, debugging, batch jobs, or exact version control:

```sh
agtx search "summarize PDFs and Word files"
agtx install web_search web_fetch
agtx run web_fetch https://example.com
```

## Manual execution fallback

Manual execution is a secondary path, not the default user experience. Use it when a human needs to verify behavior, debug a task, reproduce an agent decision, batch jobs, or pin versions.

Agents may surface these commands only when the user explicitly asks for manual control or when a workflow requires human-operated debugging.

## Capability packs

### web_search

- Use when the agent needs to discover relevant web pages, references, candidate sources, or search-result summaries.
- Inputs: natural-language query, locale, optional freshness constraints.
- Outputs: ranked results, source metadata, short evidence snippets.

### web_fetch

- Use when the agent needs to read a known URL or extract article text and metadata.
- Inputs: URL, optional session context, optional browser relay requirement.
- Outputs: title, canonical URL, main content, metadata, extraction notes.

### deep_research

- Use when the task requires multi-step evidence gathering, synthesis, product analysis, UI review, or decision support.
- Inputs: research question, scope, depth, preferred output format.
- Outputs: structured report, evidence trail, caveats, next actions.

### ocr

- Use when text must be extracted from screenshots, scans, PDF pages, UI images, or photos.
- Inputs: image or page file, optional language hints.
- Outputs: text, structure, coordinates, confidence notes.

### audio

- Use for speech recognition, speech synthesis, meeting notes, or batch audio processing.
- Inputs: audio file or text, optional language and speaker hints.
- Outputs: transcript, synthesized audio, timeline, summary notes.

### docx / xlsx / pptx / pdf

- Use for native document reading, summarization, conversion, indexing, and structured extraction.
- Inputs: document file, extraction goal, optional schema.
- Outputs: structured text, metadata, tables, summaries, extracted fields.

## Reliability model

- Native binary delivery avoids Python, NPM, and fragile runtime setup.
- Telemetry is desensitized before analysis and is used for diagnosis, upgrades, and compatibility only.
- Harness fixtures cover inputs, outputs, edge cases, and platform differences.
- Eval gates prevent low-quality versions from entering stable channels.
- Staged rollout and rollback keep upgrades reversible.