On April 15, 2026, OpenAI shipped Agents SDK v0.14.0 — the toolkit that lets developers build autonomous AI workers (agent — a program that plans multi-step tasks on its own instead of waiting for prompts). The release notes brag about 100+ supported language models via LiteLLM, a translation layer that lets one agent framework talk to many AI vendors. Scroll the list and there it is: xAI's Grok, sitting politely between Claude and Gemini. A founder reading this sees xAI as a first-class citizen in the agent stack. The compatibility matrix says yes.

The deployment roster says no.

Between April 8 and April 15, 2026, three agent platforms landed. Anthropic's Managed Agents on April 8 — Claude-only, priced at $0.08 per session-hour, with Notion, Rakuten, Asana, Sentry, and Atlassian as named early customers. Rakuten claims 10x faster agent shipping and a 22-point jump in task success over basic prompting. OpenAI's SDK on April 15 with sandboxing partners Cloudflare, Vercel, E2B, and Modal. Zed v0.232.2 the same day, adding nine models to its Pro tier. Across all three launches, the count of named enterprise customers running Grok in production: zero 😹.

This is the part that matters. "Compatible" and "chosen" are different sports.

LiteLLM compatibility is a shim — a thin adapter that rewrites OpenAI-shaped API calls (API — the plug developers use to call a model) into whatever format the vendor expects. Sounds identical on paper. In practice, function calling (the mechanism where an agent tells a tool what to do) varies wildly by model. A GitHub issue on the OpenAI Agents SDK repo — #1056 — captures the vibe: a developer asked how to wire Grok-4 in and got a community workaround (model="litellm/xai/grok-3-mini"), not official xAI docs. xAI didn't write the integration. Someone else did, for free, in a comment thread 🙀.

Now the receipts. Every April 2026 enterprise case study — Notion's workspace delegation, Rakuten's Slack agents via Claude Cowork, Sentry's automated debugging, Cisco and Nvidia inside Anthropic's Project Glasswing, Allianz's contract — names Claude or GPT. None name Grok. On benchmarks, Gemini 3.1 Pro leads SWE-bench Verified (agent coding test) at 78.80%, with Claude Opus 4.6 and GPT 5.4 tied at 78.20%. xAI has published no agent-specific tau-bench results. It has no coding CLI like Claude Code, Codex, Gemini CLI, or Jules. It has an Agent Tools API on Grok 4.1 Fast — code execution, web search, MCP support — but no managed-agent product, no reference architecture, no outcome-backed customer story 😾.

The tradeoff for buyers is nasty. "Compatible" and "supported" look identical in a vendor slide deck, then cost 2–3x more developer hours on retries, tool-schema tweaks, and missing observability (observability — the dashboards that show you why your agent failed at 3am). xAI has little incentive to close that gap. All 11 original co-founders are gone as of late March 2026. CFO Anthony Armstrong left April 10. Grok 5 missed its Q1 target and slipped to Q2. The headline moves — FedRAMP High pursuit with USDA sponsorship announced April 15, Grok Computer in private beta since April 13, XChat scheduled for April 17 — are distribution plays and consumer bets. None of them ship a Rakuten-style case study 🐈‍⬛.

So if you're picking an agent model in Q2 2026, the checklist isn't "who's on the compatibility matrix." Everyone is. The checklist is: first-party agent SDK? Named enterprise customers with published outcomes? Agent-benchmark numbers the vendor will defend? On all three, xAI is empty. Colossus 2 scaling from 1 GW to 1.5 GW in Memphis is the wrong leverage point when the bottleneck is ecosystem maturity, not GPUs.

The agent-platform shakeout is sorting labs by who shows up in case studies. Without a real SDK and named customers by Q3 2026, Grok ends the year as a checkbox, not a choice 😼.