On December 9, the Linux Foundation announced its latest project: the Agentic AI Foundation (AAIF). A few leaders in the agent space came together and created this to host donated agent projects: MCP from Anthropic, goose from Block, and AGENTS.md from OpenAI.
These projects (two of which started as internal devtools) have been open-sourced to promote transparency in a young, quickly growing space. The founding orgs are hoping that the AAIF will set standards for agent building, which will help others in the field know how to build safely and smartly.
What does this mean for developers?
It’s still a little early to tell, but if all goes according to plan, this could help improve the quality of available agentic devtools. Devs who are using agents to write code will have a few trusted (and vetted) options at hand, since best practices will be well-documented and readily available for orgs building these tools. In order to stay competitive, vendors will start to support the “standard” tools/integrations hosted by the AAIF.
TLDR: Time will tell. This could lead to more user-centric practices and less vendor lock-in for agentic devtools. At the same time, it could cement suboptimal practices as standard, which has some devs concerned.
Should I start using the AAIF incubated tools?
Being founding tools of the AAIF, it might seem like MCP, goose, and AGENTS.md are the new industry standards for building with agents. These tools are positioned for agentic tooling in the same way that OpenAPI and OCI (Open Container Initiative) are for APIs and containers.
Agents are such a fast-growing space that we think this case might look different. Researchers are still figuring out the most effective ways to build/prompt agents, and it’s highly probable that the way we use agents today is totally different from how we’ll use them 6-12 months down the line.
For example, MCP has some severe shortcomings and vulnerabilities, and thus using subagents or skills is a better option for most tasks. So yes, use MCP (e.g. it’s quite valuable for anything requiring real-time data streams), but make sure that you first understand whether a task is MCP-worthy.
The AGENTS.md standard is an obvious one, and you’re likely already using it (unless you use Claude Code, which doesn’t yet support it). If you haven’t yet implemented rules for your coding agent, you can do so quickly by creating a Markdown file and writing out your desired agent behavior.
And Block’s goose serves a very valuable role: it’s a fully open source coding agent. goose is well-loved, and can be used with dozens of LLMs. It’s definitely worth a try if you want a vendor-agnostic agent and/or one that is extremely extensible. That being said, we believe there’s still a place for Claude Code and Cursor, but that doesn’t discount how refreshing it is to see a high-quality open source alternative.
From here, it’ll be interesting to see what other tools the AAIF adopts, and whether they have a large impact on the agent landscape as a whole.
Watch this space
The launch of the AAIF is especially exciting because it’s the first time we’ve seen an industry open standard for agentic tools. There’s still a lot of uncertainty what this can mean for the agent space or for devs, but for now, we’re just watching to see how it evolves. It hosts three projects that have already been in wide rotation for agentic dev. The agent space is changing so quickly, so it’s tough to really know, and the next project(s) it adopts might not even exist yet.
