# The coding-agent harness gets standards: what the Agentic AI Foundation changes

- **Published:** 2025-12-10
- **Updated:** 2026-10-12
- **Category:** Harnesses & Tooling
- **Author:** Sinthetix
- **URL:** https://sinthetix.com/blog/agentic-ai-foundation-harness-standards

> On December 9 MCP, AGENTS.md and goose moved under a Linux Foundation fund. Why the harness around a model now matters as much as the model.

On December 9, 2025, the Linux Foundation announced the Agentic AI Foundation (AAIF) and named its three founding projects: Anthropic's Model Context Protocol (MCP), Block's goose and OpenAI's AGENTS.md [1]. The same day Anthropic said it was donating MCP to the new foundation, which it describes as a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI [2]. Coding agents, the tools that read a codebase, run commands and open pull requests, are where these three pieces are most visible today.

**Key takeaways**

- MCP, goose and AGENTS.md now sit in one vendor-neutral foundation backed by AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI at the top membership tier [1].
- The three projects cover different layers: tool access (MCP), the agent runtime (goose) and project instructions (AGENTS.md).
- Early evidence suggests the harness around a model can shift coding-benchmark results, so buyers should evaluate the whole stack, not just the model.

## What each project does

The announcement describes MCP as an open standard for connecting AI models to tools, data and applications, goose as an open-source, local-first agent framework, and AGENTS.md as a markdown format that gives coding agents project-specific guidance [1].

| Project | Origin | Layer | Reported adoption |
| --- | --- | --- | --- |
| MCP | Anthropic (open-sourced Nov 2024) | How agents reach tools and data | 10,000+ published servers [1] |
| goose | Block (early 2025) | Local agent runtime combining models, tools and MCP | Not stated |
| AGENTS.md | OpenAI (Aug 2025) | Instructions a coding agent reads in a repository | 60,000+ open-source projects and frameworks [1] |

Anthropic adds that MCP SDKs see more than 97 million monthly downloads across Python and TypeScript, and that clients include ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code [2]. These are vendor-reported numbers, but they come from the organizations that would be exposed if they were wrong.

## Why neutral governance matters

A protocol that one vendor controls is a risk for everyone else building on it. The foundation lists Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI as platinum members, with companies such as Cisco, IBM, Oracle, Salesforce, SAP and Snowflake at the gold tier [1]. Anthropic says MCP's governance model "will remain unchanged," with maintainers continuing to prioritize community input [2]. In practice that means the protocol's roadmap is meant to stay open to contributors rather than move under a single owner.

For a business, the benefit is mundane but real: an internal tool exposed once through MCP, or a repository documented once in AGENTS.md, should work with more than one vendor's agent. That reduces the cost of switching vendors later. It does not remove the work of securing those connections.

The technical direction is visible too. Anthropic says the November 25 MCP specification release added asynchronous operations, statelessness, server identity and official extensions [2]. Those are the features that matter for long-running, production deployments rather than demos.

## The harness is part of the result

The word "harness" refers to everything around the model: the loop that calls it, the tools it can use, how context is managed and the instructions it is given. A paper published a week earlier shows why that matters. DeepSeek's V3.2 report gives a score of 46.4 on Terminal Bench 2.0 and states it was obtained with the Claude Code framework, because the team's context management for thinking mode was incompatible with the benchmark's standard Terminus harness. With Terminus in non-thinking mode, the same model scored 39.3 [3]. The two runs differ in mode as well as harness, so the gap is not a clean measure of harness effect. For SWE-bench Verified, the report's primary figure of 73.1 came from an internal framework, with results from the Claude Code and RooCode frameworks and non-thinking mode ranging from 72 to 74 [3].

**DeepSeek-V3.2 on Terminal Bench 2.0, by harness and mode**

Accuracy as reported by the model's authors

Unit: %

| Category | Terminal Bench 2.0 |
| --- | --- |
| Claude Code framework (thinking) | 46.4% |
| Terminus (non-thinking) | 39.3% |

*The two runs differ in both harness and thinking mode; the gap is indicative, not a controlled comparison. Source: [3], DeepSeek-AI (arXiv:2512.02556)*

The SWE-bench numbers say the opposite: across several frameworks, the spread was small [3]. The honest reading is that harness sensitivity varies by task and benchmark, which is itself a reason to test on your own work.

## What to watch next

- **Security.** Standardizing how agents reach tools also standardizes the attack surface. Every MCP server an agent can call is a permission to review.
- **Convergence.** Foundation membership is not a guarantee that vendors will implement the specification identically. Compatibility will be proven by clients and servers working together, not by press releases.
- **Events and process.** The foundation notes the next MCP Dev Summit is planned for New York City on April 2 and 3, 2026 [1].

## Implication

If your engineering teams are adopting coding agents, write project instructions in AGENTS.md, expose internal tools through MCP rather than one vendor's proprietary plugin format, and evaluate any agent as a whole system on your own repositories.

## Sources

1. [Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF), Anchored by New Project Contributions Including Model Context Protocol (MCP), goose and AGENTS.md](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation), The Linux Foundation, 2025-12-09
2. [Donating the Model Context Protocol and establishing the Agentic AI Foundation](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation), Anthropic, 2025-12-09
3. [DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models](https://arxiv.org/abs/2512.02556), DeepSeek-AI (arXiv:2512.02556), 2025-12-02

## How Sinthetix helps

**AI Agents & Automation**: Intelligent agents and automated workflows that reduce manual work and streamline operations.

- Multi-agent systems
- Task automation
- AI-driven decision logic

[Explore AI Agents & Automation](https://sinthetix.com/services/ai-agents-automation) | [Talk to us](https://sinthetix.com/contact)

Sinthetix is an AI engineering studio. We ship agents, automation, and internal tools for operators who need results, not reports. [About Sinthetix](https://sinthetix.com/about)