AI’s next race: agents

The new contest in artificial intelligence is not who writes the prettiest paragraph. It is who can take a messy job like “reconcile invoices, check the contract, email the vendor, and update the spreadsheet” and keep going without a human nudging every step. (openai.com) That shift is showing up in the products. OpenAI says its Responses Application Programming Interface is meant for “agentic applications” that use built-in tools like web search, file search, and computer use inside one workflow instead of one chat reply at a time. (openai.com) An agent is basically a language model with hands. The model reads the request, chooses a tool, takes an action, checks the result, and then decides the next action the way a person might bounce between a browser tab, a file folder, and a form. (developers.openai.com) That is why the word “orchestration” keeps coming up. Orchestration is the software layer that decides which model to call, which tool to use, what memory to keep, and when to stop, like a project manager routing work across a team instead of doing the work alone. (github.com) The other word is “duration.” Anthropic says Claude Opus 4.6 was built to “sustain agentic tasks for longer,” and added context compaction so the system can summarize its own working memory and keep operating without hitting context limits. (anthropic.com) That sounds abstract until you compare it with ordinary chat. A chatbot that answers one question well can still fail on a 40-minute task if step 17 breaks, the login expires, or the model forgets why it opened the third spreadsheet in the first place. (anthropic.com) The big platforms are now selling exactly that longer loop. Microsoft describes autonomous agents in Copilot Studio as systems that monitor data, react to triggers, and run workflows in the background without waiting for a new prompt each time. (learn.microsoft.com) Google is pitching the same idea from the enterprise side. Google Agentspace is designed to combine search, reasoning, and company data so employees can ask for a result and the system can plan, research, generate content, and take actions from one request. (cloud.google.com) This is why old leaderboards are starting to look incomplete. A model can top a writing benchmark and still be worse for real work than a slightly weaker model wrapped in better memory, safer tool permissions, and a tighter retry loop. (openai.com) For companies buying this stuff, the useful test is not “write me a marketing email.” The useful test is “take a support ticket from inbox to refund,” or “take a sales lead from web form to customer record,” and then measure completion rate, error rate, handoff rate, and time saved. (learn.microsoft.com) That is where the next race is heading. The winners may not be the models that sound smartest in a demo, but the systems that can stay on task the longest, recover when a tool fails, and finish the whole job before a human notices there was work to do. (anthropic.com)