Every Big Vendor Just Made the Same Bet: Agents That Run for Days

OpenAI, Alibaba, Google, and Anthropic all shipped agents this week measured in hours and days, not prompts. The race quietly shifted from how smart an agent is to how long it can run unattended — and that changes how you evaluate them.

The question about coding agents used to be how smart they were in a single turn. This week, the major vendors quietly changed it to how long they can run without you watching. In the span of a few days, OpenAI, Alibaba, Google, and Anthropic each shipped or detailed agents measured in hours and days rather than prompts — and that reframes what "good" even means.

OpenAI set the tone with its Codex Thursday launches, introducing Goal mode: you set a goal "that it can work towards for hours or even days," now live in the app, IDE extension, and CLI. The same week, VentureBeat reported that Alibaba's Qwen3.7-Max can run autonomously for 35 hours and plugs into external harnesses like Claude Code, putting a Chinese lab squarely in the same long-horizon territory. Google's entry, reported by AI Automation Global from Google I/O 2026, is Gemini Spark, a 24/7 agent on dedicated Cloud VMs that keeps working while you're offline.

Anthropic pushed duration toward a different end. Its Project Glasswing update reports the initiative and its partners have found more than 10,000 high- or critical-severity vulnerabilities since its launch last month, work it attributes to models like Claude Mythos Preview. The number isn't the real story — the warning is. Anthropic says the software industry "will need to adapt to the volume of vulnerabilities" autonomous models can surface, which is a vendor telling customers the bottleneck is about to move from finding bugs to fixing them.

For builders, the practical takeaway is that orchestration now matters more than picking a single model. Simon Willison's Datasette Agent alpha runs against local Gemma 4 and Qwen 3.5/3.6, and the open-source camp — see OpenHands on dev.to — is betting on a model-agnostic stack you reroute from a single config file. When an agent runs for a day, choosing wrong compounds for a day, so a routing layer you own is a hedge, not a purity test.

For buyers, a date-stamped caution: every duration figure above is a vendor or single-report claim as of May 2026, not an independent benchmark. "Runs for 35 hours" is not the same as "produces 35 useful hours," and no one has published the latter yet.

The bottom line: Stop evaluating these tools on a single clever answer and start testing how long they hold a goal before they drift — that's the axis every vendor is now competing on.

Sources