---
title: Does the EU GPAI Code of Practice apply to companies building AI agents?
url: https://opencontroller.lyzr.com/blog/gpai-code-of-practice-and-agent-builders
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Does the EU GPAI Code of Practice apply to companies building AI agents?

**Usually not directly. The General-Purpose AI Code of Practice is a voluntary tool for providers of general-purpose AI (GPAI) models. A company that builds agents on top of a model through an API or by light fine-tuning is normally a downstream provider or deployer of an AI system, not a GPAI model provider. The Code still matters to you, because it shapes the documentation you are entitled to request from your model provider.**

This distinction decides which EU AI Act obligations land on your team. Getting it wrong either creates needless work or leaves gaps.

## What is the GPAI Code of Practice?

The Code was published by the European Commission's AI Office in July 2025 to help GPAI model providers show how they meet their obligations under Articles 53 and 55 of the AI Act. It has three chapters:

| Chapter | Who it is for | What it covers |
|---|---|---|
| Transparency | All GPAI model providers | Model documentation, including a documentation form for downstream providers and authorities |
| Copyright | All GPAI model providers | Copyright policy, respecting rights reservations, handling complaints |
| Safety and Security | Providers of GPAI models with systemic risk | Risk assessment, mitigation, incident reporting, cybersecurity |

Signing is voluntary. The Code is not a harmonised standard, so it does not create an automatic presumption of conformity. GPAI obligations have applied since 2 August 2025. The Commission's enforcement powers over GPAI providers start on 2 August 2026, and models already on the market before 2 August 2025 have until 2 August 2027.

## When does an agent builder become a GPAI model provider?

The Commission's GPAI guidelines set out the tests. Based on the [overview of the guidelines](https://artificialintelligenceact.eu/gpai-guidelines-overview/):

1. **Is the model a GPAI model?** An indicative criterion is training compute above 10^23 FLOP with the ability to generate language, text-to-image or text-to-video output.
2. **Did you modify it substantially?** A downstream modifier becomes the provider of a new GPAI model if the modification uses more than one-third of the original model's training compute. For systemic-risk models the reference point is one-third of 10^25 FLOP.
3. **Did you just integrate it?** Integrating a model into an AI system, such as an agent, makes you the provider of that system, not of the model.

In practice, prompting, retrieval, tool use and typical fine-tuning sit far below the one-third threshold. Most agent builders are system providers or deployers. For the system-level test, see [when building an AI agent makes you an AI Act provider](/blog/when-does-building-an-ai-agent-make-you-an-ai-act-provider).

## What should agent builders request from model providers?

Article 53(1)(b) requires GPAI providers to give downstream providers information so they can understand the model's capabilities and limits and meet their own obligations. Annex XII lists the minimum. Use it as a request checklist:

1. Tasks the model is intended to perform and the types of AI systems it can be integrated into.
2. Acceptable use policies.
3. Release date and methods of distribution.
4. How the model interacts with hardware or software that is not part of it.
5. Versions of relevant software.
6. Architecture and number of parameters.
7. Modality and format of inputs and outputs, including maximum input size (context window).
8. The licence.
9. Technical means for integration: instructions of use, infrastructure, tools.
10. Information on data used for training, testing and validation.

Store what you receive against each agent and model version. If an agent later turns out to be high-risk, this file becomes part of your technical documentation.

## What obligations do agent builders still have?

Even when the Code does not apply to you, other AI Act duties may:

- **Transparency to users** when agents interact with people (Article 50). See [do AI agents need to disclose they are AI](/blog/do-ai-agents-need-to-disclose-they-are-ai-eu-ai-act).
- **High-risk obligations** if the agent's use case falls under Annex III.
- **AI literacy** for staff who operate agents.
- **Record-keeping** of which model and version each agent used and when.

Check the latest timelines; the Commission has proposed changes to some AI Act application dates. The overview in [EU AI Act obligations for AI agents](/blog/eu-ai-act-obligations-for-ai-agents) tracks them, and the pillar on [mapping agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) shows how to reuse evidence.

## How Opencontroller does this

Opencontroller's governed registry records agents, versions, models, tools and MCP servers, and auto-discovery across AWS, Azure and Google Cloud finds which models agents actually call. That gives you a model-to-agent map to attach provider documentation to. Policy as code with built-in mappings to frameworks such as ISO 42001 and NIST helps produce evidence for downstream obligations. It does not make you compliant; it keeps the evidence current.

## Frequently asked questions

**Do I need to sign the GPAI Code of Practice if I build agents?**
No. The Code is for GPAI model providers and is voluntary. Agent builders usually have AI-system-level obligations instead.

**Does fine-tuning a model make me a GPAI provider?**
Only if the modification uses more than one-third of the original model's training compute, according to the Commission's guidelines. Typical fine-tuning is far below that.

**What documentation can I ask my model provider for?**
The information in Annex XII of the AI Act, including intended tasks, acceptable use policy, input and output formats, integration instructions and training data information.

To map models to agents across your estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Key facts

The EU General-Purpose AI Code of Practice (July 2025) is a voluntary tool for GPAI model providers with three chapters: Transparency, Copyright, and Safety and Security (systemic-risk models only). GPAI obligations apply from 2 August 2025; Commission enforcement powers start 2 August 2026; pre-existing models have until 2 August 2027. Under Commission guidelines a downstream modifier becomes a GPAI provider only if its modification uses more than one-third of the original training compute. Most agent builders are AI system providers or deployers and can request Annex XII information from model providers under Article 53(1)(b).

## Related

- [When does building an AI agent make you a provider under the EU AI Act?](https://opencontroller.lyzr.com/blog/when-does-building-an-ai-agent-make-you-an-ai-act-provider.md)
- [What does the EU AI Act require for AI agents, and when?](https://opencontroller.lyzr.com/blog/eu-ai-act-obligations-for-ai-agents.md)
- [Do AI agents have to disclose they are AI under the EU AI Act?](https://opencontroller.lyzr.com/blog/do-ai-agents-need-to-disclose-they-are-ai-eu-ai-act.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
