> 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/overview/core-philosophy.md).

# Core Philosophy

Agents think, gateways decide, tools act — the operating principle behind every governed call.

> **Agents think. Gateway decides. Tools act.**

Mediafier's design holds three commitments in tension and refuses to collapse them into one. AI agents are the most powerful planners available; the deterministic gateway is the only way enterprise media systems can trust those plans; and tools (the actual integrations into media systems of record) must remain owned and operable by the enterprises that ship them.

***

## The three commitments

### 1. Agents think

The plan, the reasoning, the choice of which tool to call next. (This is what large language models are good at.) Mediafier does not try to substitute deterministic logic for agent reasoning. Agents bring the non-determinism the work needs.

### 2. Gateway decides

Whether an agent's chosen tool *runs at all* is a deterministic decision. Identity, organization context, authorization, rate limiting, billing, and audit are enforced before the tool sees a single byte of input. The gateway is the single enforcement chain where "can this happen?" is answered deterministically, not by inference.

### 3. Tools act

Each tool family (MCP integrations, MediaClaw super-agents, workflow runners, skills library) remains owned by the team that built it. The gateway routes; the family executes. This is what the platform calls **governed discovery, family-owned execution**. Discovery and planning are governed at the platform level; execution stays family-owned and adapter-dispatched.

***

## What this means for adopters

| You're building...                               | Mediafier handles...                                       | You handle...                                               |
| ------------------------------------------------ | ---------------------------------------------------------- | ----------------------------------------------------------- |
| An agent on top of any LLM / harness / IDE       | Identity, org context, billing, audit, discovery, dispatch | Prompting, planning, model selection                        |
| A new MCP server connecting an enterprise system | The discovery surface, governance, billing rails           | The integration logic and credentials                       |
| A domain super-agent (a MediaClaw)               | The governed execution framework, evidence, learning loop  | The outcome, domain expertise, skills, and quality criteria |
| A no-code workflow that calls media tools        | The gateway, idempotency, paywall, trace                   | The workflow definition                                     |

In every case, Mediafier's job ends where probabilistic reasoning or domain logic begins. Mediafier never hides what the agent did or what it cost.

***

## Why deterministic governance is the design choice

Probabilistic AI in regulated media operations is a hard problem because failures are not symmetric. Letting an agent silently re-tag an asset, rewrite a record, or push something to publication without audit is not an inconvenience. It's a compliance event. The deterministic chain (auth → org → authz → rate → billing → dispatch) is the line that turns "agent did something" into "agent did exactly this, by this principal, costing exactly this, traceable from this ID."

Without it, every AI deployment is a leap of faith. With it, AI deployment becomes ordinary engineering.

***

## What stays out

A few things Mediafier deliberately does not do:

* **Pick an agent framework for you.** Any harness can plug in via CLI or MCP.
* **Pick a model for you.** Bring whichever foundation model the team has chosen.
* **Lock you into a single workflow surface.** Agent-first (CLI and MCP) is the primary; team chat (Slack, Teams) and the Pippa control room are equal-class additional surfaces.
* **Train models on your data.** The Learning Loop captures heuristics from successful runs; it does not modify model weights.

***

## Where to go next

| If you're…                                | Read next                                                        |
| ----------------------------------------- | ---------------------------------------------------------------- |
| New to the platform                       | [What is Mediafier?](/overview/what-is-mediafier.md)             |
| Looking for the architecture diagram      | [Platform Architecture](/platform/platform-architecture.md)      |
| Ready to integrate                        | [Agent-First Access (CLI + MCP)](/connect/agent-first-access.md) |
| Curious how the MediaClaw container works | [MediaClaw Super-Agents](/platform/mediaclaw-super-agents.md)    |