Every GTM engineer now builds on someone else’s agent runtime — and 2026 has delivered a controlled experiment in what that dependency costs. In one seven-month window, OpenAI launched and then deprecated its flagship visual agent builder, while the Model Context Protocol moved from single-vendor stewardship to a neutral foundation backed by every major lab and cloud. The lesson is not “avoid OpenAI” or “bet on Anthropic.” It is that the protocol layer and the product layer are aging at completely different speeds — and your GTM stack should be attached to the slow-aging one.

The seven-month product

AgentKit launched at OpenAI’s DevDay on October 6, 2025: a visual Agent Builder canvas, the embeddable ChatKit, a Connector Registry, and Evals — bundled into standard API pricing with no platform fee (announcement). On June 3, 2026, Agent Builder and Evals were deprecated, with full shutdown scheduled for November 30 and migration pointed at the code-first Agents SDK or ChatGPT Workspace Agents; ChatKit survives (deprecations page). Seven months, launch to deprecation. Any GTM team that built its routing flows, SDR-agent logic, or enrichment orchestration in that canvas now owns a rewrite with a hard deadline.

The protocol went the other way

In the same window, the plumbing consolidated toward neutrality. In December 2025, Anthropic, OpenAI, and Block co-founded the Agentic AI Foundation under the Linux Foundation — Anthropic contributing MCP, OpenAI contributing the AGENTS.md spec, Block contributing goose — with AWS, Google, Microsoft, Cloudflare, and Bloomberg as platinum members (Linux Foundation). MCP’s public server count passed 9,400 before the handoff. A protocol governed by a foundation whose members include every lab and every cloud does not get deprecated in seven months; it gets boring, which is exactly what infrastructure should be.

What this means for a GTM stack specifically

GTM tooling is unusually exposed to this dynamic because the stack is a chain of integrations by nature: signals feed enrichment, enrichment feeds sequencing, everything reads and writes the CRM. Each link is a place where you either wired to a vendor’s proprietary surface or to a protocol. The 2026 platform consolidation — three signal vendors absorbed into CRM platforms in four months — raises the stakes: when your signals vendor becomes a HubSpot feature, an MCP-shaped integration migrates; a bespoke one dies with the standalone product. The same logic applies to the agent runtime itself: n8n workflows calling models through standard APIs survived AgentKit’s deprecation without noticing it happened.

The build rules

1. Logic in code or open workflow tools, not vendor canvases. The Agents SDK, n8n, or plain scripts age at your pace; visual builders age at the vendor’s. 2. Integrations through MCP where a server exists — and prefer vendors that ship one, because it is now a proxy for how seriously they take portability. 3. Model-agnostic by default: route through an abstraction so that swapping Claude, GPT, or Gemini is a config change, not a rewrite — the FM layer is repricing too fast to marry. 4. Treat first-party GUIs as demos. Use them to prototype; rebuild anything load-bearing on the SDK layer before it carries revenue.

The uncomfortable caveat

Protocols are necessary but not sufficient. MCP standardizes how an agent reaches your CRM; it does not stop the CRM owner from bundling a competing agent, and it does not protect you when the vendor behind an MCP server gets acquired and sunsets the product — the connector outliving the tool is cold comfort. The full defensive posture is protocol-first plumbing plus the contract terms we keep repeating this year: data portability and change-of-control clauses on every vendor that holds your revenue data. Our best-practices frameworks and the GTM engineer’s toolkit carry the working checklists.