The agent graveyard: what died in two years
The agent era is barely two years old and it already has a respectable cemetery. Not failed startups — that's ordinary attrition. These were flagship surfaces from the biggest AI companies on earth, launched with keynotes, dead within months. The pattern in how they died says more about where agent interfaces are going than most launch posts do.
Here's the register, dates and all.
The plots
OpenAI Operator — seven months. Launched January 23, 2025 as a standalone computer-use agent product; merged into ChatGPT agent on July 17, 2025; the standalone site shut down that August. The capability — an agent operating a browser on your behalf — didn't die at all. It moved into the product people already had open.
ChatGPT Atlas — under ten months. OpenAI's standalone agentic browser, launched October 21, 2025, macOS only. Discontinuation announced July 9, 2026; it stops working August 9 — folded into the ChatGPT desktop app and a Chrome extension. Windows and mobile versions never shipped at all.
AgentKit's Agent Builder — eight months to a deprecation notice. The visual agent-building suite announced at DevDay in October 2025; by June 3, 2026 the deprecations page had scheduled the visual builder and its evals for a November 2026 sunset, pointing builders back to the code-first SDK. Only ChatKit survives.
IBM's ACP — roughly five months. The Agent Communication Protocol launched in March 2025 as a rival in the agent-to-agent race and merged into A2A that August, repo archived with a banner saying so. Its acronym was promptly recycled by Zed's unrelated Agent Client Protocol, which is a genuine hazard for anyone reading 2025 coverage today.
Agent Protocol — dormant. The AI Engineer Foundation's early attempt at a common agent API predates the current wave and has been quiet since spring 2025. First to the layer, adopted by no one that mattered.
And one that didn't die but stopped being itself: Dia, The Browser Company's agentic browser, was acquired by Atlassian for roughly $610 million in September 2025 — an exit, not a failure, but another standalone surface that ended up inside something bigger.
What actually kills them
Line the graves up and three patterns fall out.
Capabilities survive; standalone surfaces don't. Operator's computer use lives on inside ChatGPT. Atlas's browsing lives on inside the desktop app and a Chrome extension. Nothing about the ability failed — what failed, twice, at the same company, within a year, was the bet that a new agent capability deserves its own destination app. Distribution gravity won both times. The lesson generalizes: agents go where users already are, which is why the in-chat app surfaces and browser extensions keep growing while standalone agent products keep getting absorbed.
Visual agent-building keeps losing to code. Agent Builder is the highest-profile casualty of a pattern anyone who watched the low-code wave will recognize: the demo is spectacular, the maintenance story isn't. Agents are software; the people responsible for software that acts autonomously want version control, tests, and diffs — not canvases. Eight months from keynote to deprecation notice is a fast admission.
Protocol races consolidate early now. ACP merging into A2A five months after launch — and AP2 and x402 each landing in neutral foundations within months of shipping — suggests the industry learned something from the decade browser vendors spent fragmenting standards. Nobody wants to be the loser of a protocol war, so the wars are ending before they start. The graveyard's protocol section may stay small; its product-surface section will not.
Why we keep this list
Every claim about this space rots fast — the average lifespan in this article is well under a year. That's the practical reason this site's tracker treats deaths as first-class entries with dates and sources: knowing that "just use Operator" or "build it in Agent Builder" is dead advice is worth as much as knowing what to use instead. Date every claim you read about agents, including ours. The stones here all have two dates on them, and the second one arrived quicker than anyone's roadmap said it would.