> For the complete documentation index, see [llms.txt](https://docs.mediafier.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.mediafier.ai/platform/mediaclaw-super-agents.md).

# MediaClaw Super-Agents

MediaClaws: bounded, observable expert agents with a Cognitive Rig, Evidence Cards, human-controlled skill promotion, and mastery scorecards.

> **A deterministic container around a probabilistic brain.** A MediaClaw is what happens when an LLM-powered agent gets the structure of a real production worker — a structured operating loop, an evidence trail, governed skills, and a learning loop that compounds run over run.

A MediaClaw runs domain-specific media tools and gets better at them each week. Work that used to take a coordinated multi-person handoff completes in seconds, with the operator's judgment retained at the points that matter.

***

## What's inside

Four mechanisms work together:

### 1. Cognitive Rig

A Cognitive Rig is the structured **Plan → Execute → Validate → Learn** operating loop that makes a MediaClaw bounded, observable, and repeatable, recording the evidence needed to inspect how a run progressed.

### 2. Evidence Cards

One **immutable record** per step. Every step writes a card capturing what ran, what it produced, and the trace ID that joins it back to the gateway and audit chain; usage is captured for billing attribution and settlement. Cards are append-only and immutable — enforced at the platform's data layer, not by convention. When a stakeholder asks "why did the agent do that?", you open the cards.

### 3. Skills Governance

A reusable skill is a tool-using competency a MediaClaw can deploy. New and revised skills are validated against the current approved skill before they can be used, and require **human approval** before activation. The authenticated product shows each skill's review state.

Promotion is **human-controlled**. The Learning Loop can propose a promotion; it cannot execute one.

### 4. Learning Loop

Every successful run is a candidate for **heuristic extraction**. The Learning Loop reads the run's evidence trail, looks for patterns worth promoting, and drafts a proposal — a refined skill, a new heuristic, an updated operating directive. Proposals enter governance as pending proposals and are not deployed without human approval.

The Learning Loop **does not modify model weights**. There is no fine-tuning. The model is whichever foundation model the team has chosen; the Learning Loop only changes the skill library and the cognitive-rig directives that wrap it.

***

## How MediaClaws compose

A MediaClaw can call another MediaClaw. A Podcast Ingest claw can call an Asset QC claw which can call a Brand Safety claw — each one its own governed, billable surface. Composition does not collapse governance: every cross-claw call goes through the gateway, gets a trace ID, produces evidence, and is metered against the calling organization's credits.

Composition is how a single agent task that used to take a half hour across three operators becomes a half-minute across three claw calls.

***

## What you commit to when you ship a MediaClaw

A MediaClaw is not a one-shot script. Shipping one is a commitment to:

* A **Cognitive Rig definition** — the operating loop's shape and the validation criteria.
* A **skills set** — versioned and governed, with promotions reviewed by humans.
* An **evidence contract** — every step writes a card; cards are append-only.
* A **billing surface** — every step is metered through the gateway.

In return, the platform gives the claw governed tool access, durable context, inspectable runs, and learning over time.

***

## What MediaClaws are not

* They are not a generic LLM chat. The container enforces structure; every action goes through a defined stage of the operating loop and a governed tool.
* They are not unrestricted. Every tool call passes through the gateway, and the cognitive rig only reaches tools the platform has governed.
* They are not unmanaged learners. The Learning Loop's outputs are proposals, not actions. Humans approve.

***

## Where to go next

| If you're…                                              | Read next                                                                           |
| ------------------------------------------------------- | ----------------------------------------------------------------------------------- |
| New to the platform                                     | [What is Mediafier?](/overview/what-is-mediafier.md)                                |
| Studying decision memory and the learning loop in depth | [Decision Tracing & Learning Loop](/platform/decision-tracing-and-learning-loop.md) |
| Looking at how MediaClaw surfaces are discovered        | [Governed Resources & Discovery](/platform/governed-resource-substrate.md)          |
| Looking for the gateway internals that govern claw runs | [The Gateway](/platform/gateway.md)                                                 |