Spark
By Jason Kumpf, Strategy Advisor · October 28, 2026 · 5 min read
Anthropic, NVIDIA and Google spent 2026 shipping AI agents that plan, act and check their own work. The shift opens a fresh layer of opportunity for founders who move early.
In 2026 the biggest names in AI moved from assistants that answer questions to agents that finish work. Anthropic, NVIDIA and Google each shipped tools this year that let software plan, act and check its own output over hours rather than seconds.
For founders, that shift opens a new layer of company building. The models are capable, the standards are open and the playbook is still being written. That is exactly the moment when small teams with a clear idea tend to move fastest.
Anthropic introduced Claude Opus 5.5 on September 22, 2026, and designed it for long, autonomous work. It can run for extended periods, delegate to subagents and review its own output before handing work back. It is available in Claude Code, in the Claude apps and through the Claude API.
The early results show what "long-running" means in practice. Anthropic reports that one team completed a 680,000-line code migration in under a day.
The pattern matters more than any single benchmark: work that once required a project plan and a team can now start with a clear brief and a capable agent.
Claude Code began as a command-line tool for developers, and 2026 showed how far that idea can travel. In December 2025, engineers at NASA's Jet Propulsion Laboratory ran Claude in Claude Code to plan a roughly 400-meter route for the Perseverance rover through a field of rocks on Mars.
The engineers first supplied years of their own driving data and experience. Claude then studied overhead images, laid a breadcrumb trail of waypoints in 10-meter segments and refined its plan through self-critique. A simulation that models more than 500,000 variables confirmed the route needed only minor changes, and the rover drove it successfully. Engineers estimate the approach could cut route-planning time in half.
Anthropic published the story in January 2026 as the first AI-planned drive on Mars. It captures the central lesson of the agent era. Agents perform best when experts supply context, and experts move faster when agents handle the detailed execution.
On October 22, 2026, Anthropic donated the Agent Skills standard to the Agentic AI Foundation, a Linux Foundation project that Anthropic co-founded with Block and OpenAI. A skill is simply a folder: a short instruction file plus any scripts or reference material that teach an agent one specific task. Anthropic developed the idea while building Claude Code.
Adoption has been remarkably fast. More than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, have integrated Agent Skills since the standard was published in December 2025. Developers have published 1 million skills, and installs total nearly 280 million. Engineers from Anthropic, Google, OpenAI and other companies will maintain the standard together.
For founders, open standards widen the path to customers. A well-built skill or connector can reach users across dozens of agent platforms at once, without a separate integration for each one.
At GTC in San Jose on March 16, 2026, NVIDIA introduced NemoClaw, a software stack that installs NVIDIA Nemotron open models and the new OpenShell runtime with a single command. OpenShell gives each agent its own sandbox with policy-based privacy controls, and a privacy router lets agents combine local open models with frontier models in the cloud. Jensen Huang called the OpenClaw platform "the operating system for personal AI."
NVIDIA kept building at GTC Berlin, held October 20 to 22, 2026. Highlights for agent builders included:
The message is consistent from data center to laptop. Agents will run everywhere, around the clock, close to the data and people they serve.
At Google I/O on May 19, 2026, Google declared "the agentic Gemini era." Gemini 3.5 Flash, the first model in its newest family, pairs strong reasoning with the ability to act. Antigravity 2.0 arrived as a standalone desktop app for orchestrating multiple agents in parallel, with dynamic sub-agents, scheduled background tasks and integrations with AI Studio, Android and Firebase.
Google also added Managed Agents to the Gemini API. A single API call starts an agent that reasons, uses tools and runs code in an isolated Linux environment that keeps its files and state between interactions. Developers can shape each agent with Markdown-based instructions and skills.
Across three of the most important companies in AI, the same building blocks keep appearing: long-running agents, reusable skills and secure sandboxes. When the industry converges this clearly, founders can build with confidence.
Three opportunity zones stand out as the agent stack matures:
Each zone rewards domain knowledge as much as engineering talent. The founders who stand out will pair a sharp understanding of one customer's day with the new agent toolkits.
Start small and specific. Pick one recurring workflow in your company, write down how your best expert handles it and turn that knowledge into an agent brief or a skill. Track the hours it returns to your team, then expand to the next workflow.
Then compare notes with other builders. The fastest learning in the agent era happens in conversations among founders, operators and investors who share what worked last week. The people who trade those lessons early will be the first to see the next opening.
