Building the Foundation for the Agentic AI Era

The actual news is a governance move. OpenAI, Anthropic, Google, and Block have placed their agent protocols (MCP, A2A, agents.md, Goose, and Agent Gateway) under a neutral Linux Foundation body called the Agentic AI Foundation. Angie Jones, who ran AI adoption across Block's 3,500 engineers, is a key voice explaining how the stack fits together. The question for anyone shipping agents into production: does this change what you build on?

Probably not yet. The foundation is eight months old and hasn't resolved the EU AI Act compliance question its own working group is sitting on. Standards bodies historically ratify what the market leader already shipped. The same four labs that just donated their protocols are also running heated public disputes over model launches, and neutral governance has never stopped a market leader from adding proprietary extensions that everyone ends up supporting anyway.

The more actionable thing Jones said isn't about governance at all. Developer velocity at Block didn't move until roughly November 2025, around the Claude Sonnet 4.5 release. That's a firsthand enterprise report from a real production environment, not a benchmark. It matters because it dates the moment when "AI coding agents don't help on big mono-repos" flipped from true to stale. If your internal eval of coding agents predates that window, you're working from bad numbers.

On the practical side, the open-source configuration here is genuinely unusual. MCP, A2A, agents.md, and Goose are all open, under neutral governance, with all four major contributors at the table. The interoperability layer of the agent stack isn't owned by one lab. Goose ran across those 3,500 Block engineers in production. For teams that don't want to marry a single vendor framework, building on these protocols is the low-lock-in path. Just don't confuse open rails with open models. Claude, GPT, and Gemini are still running on the tracks.

If you're wiring something up now: MCP for tool access, agents.md checked into the repo. That last part is the operational lesson worth stealing. Block didn't train 3,500 developers on agent norms. They embedded context engineering into the repos so agents behaved correctly without anyone reading a wiki. You can copy that pattern this week.

Agent Gateway deserves attention for a different reason. Base protocols ship without observability, access control, or audit logs. The Gateway is where that gap shows up, and it's the layer that breaks first at scale and that enterprise procurement will ask for before they sign. The foundation hasn't shipped the compliance piece yet, so that gap gets filled by proprietary vendors in the meantime.

One cost warning Jones glosses over: every A2A hop is a model generating tokens to coordinate another model generating tokens. Multi-agent fan-out is a cost multiplier, and the protocols make that fan-out easy to trigger and easy to under-price. Run your own token-cost model against your actual fan-out before you believe the "millions of coordinated agents" framing.

The call: By end of 2026, MCP will still be the dominant agent-tool protocol in production, but at least one founding lab will have shipped a proprietary agent capability that its top model supports and the neutral standard doesn't yet cover. Medium confidence. Labs collaborate on plumbing and compete on features, and that split always resolves toward proprietary extensions. The foundation neutralizes the plumbing layer. It doesn't neutralize the incentive to win.

The story here is about governance rather than capability. OpenAI, Anthropic, Google, and Block have placed their agent plumbing under a neutral Linux Foundation body called the Agentic AI Foundation. The standards govern how AI agents talk to tools and to each other. Angie Jones, who ran AI adoption across Block's 3,500 engineers, explains what that means in practice. The question for any team building with agents: does this change what you build on, or is it a press release with a foundation logo?

What's actually being decided: whether to build on these shared standards now. Adopting them is a bet you can walk back. The harder-to-undo risk arrives later, if "standards compliance" becomes a checkbox in big-company purchasing contracts.

The Skeptic. Standards bodies are where fast-moving things go to slow down. Jones argues this one differs because the members are paid staff with a mandate. Maybe. But the same four companies that donated the standards are still fierce rivals, and neutral governance has never stopped a market leader from shipping its own extensions that everyone else must support anyway. An eight-month-old foundation with its main compliance question unanswered is a promising committee. Picture a wall socket everyone has agreed to draw before anyone has agreed on the voltage.

The Researcher. Jones's date matters more than the announcement: developer speed at Block didn't move until roughly November 2025, when a new generation of models arrived. That is a firsthand report from a 3,500-engineer company, and it marks when "AI coding tools don't help on big codebases" went from true to out of date. If your company evaluated these tools before last November, that evaluation is stale.

The Open-Source Advocate. The standards are genuinely open, and that is real progress. Keep one eye open all the same: the models running behind those open standards still belong to the same few companies, and agent-to-agent conversations multiply usage, which means bills that grow faster than you'd guess.