OpenAI spent the second week of September renaming the thing your buyers do at work. GPT-6 Astra arrived on September 9 billed as the next generation in intelligence for work, and the Agents API followed a day later in public beta: build a cloud agent, run it on OpenAI's Codex harness, and let OpenAI run the orchestration and session plumbing underneath. Put the ChatGPT Agents API and GPT-6 Astra news together and the direction is unambiguous. Work search is becoming agentic, which means the searching increasingly happens inside a delegated task instead of a search box.
What Actually Shipped
Three concrete things, in one week.
GPT-6 Astra, the work-branded flagship. The most telling detail is OpenAI's own developer advice: get more out of Astra by revisiting your skills, your AGENTS.md files, and your task prompts, with specific triggers and a defined picture of done. That's guidance for steering a worker, and it drew twelve thousand likes from developers who understood exactly what it implied.
The Agents API in public beta. OpenAI runs the agent infrastructure; developers ship the agent. Long-running sessions and managed context mean agents that work for hours, not chat turns.
Go from idea to a working agent faster with the Agents API. Build and run cloud agents with the Codex harness, fully managed by OpenAI. We handle orchestration, long-running sessions, and context management. You focus on what makes your agent unique. Available in public beta. https://t.co/74siIaFdg4
OpenAI Developers@OpenAIDevsSep 10, 2026
The ecosystem push. A Product Hunt challenge for Astra-built projects within the week, and voice getting serious: OpenAI reports its GPT-Live-1 voice model paired with Astra completes 83.6% of customer-support tasks first-attempt on the Tau3 benchmark, against 45.7% for the previous generation. Voice agents that finish tasks are agents your customers will meet.
Where Discovery Goes When Search Becomes a Task
Now the marketing half. When an employee asks a search engine "best data pipeline tools," you can at least compete for the answer they read. When that employee tells an Astra-powered agent to evaluate pipeline vendors, produce a comparison, and recommend two, every discovery moment happens inside the run: which sources the agent trusts, which pages it can even load, which docs it can parse, which vendors survive its verification pass. The human sees the deliverable, and the deliverable has already decided who's in it.
This is the same trajectory we tracked when GPT-5.6 flipped ChatGPT toward trust-first retrieval, extended one step: fewer, more deliberate reads, chosen by source identity. Early third-party trackers already claim Astra searches substantially less than its predecessor and reads brand sources directly. Treat those numbers as provisional; treat the direction as established.
I realize I glossed over something, though. "Work search" was always a means to finished tasks, with search as the best tool available at the time. Agents collapse that distinction, which is why this shift will run faster than the last one: nobody loves searching, and everybody loves delegating.
15 hours a week manually. Or 15 minutes with RankControl.
Track citations, monitor competitors, and fix content gaps across every AI search engine. Automatically.
The Economics Say This Is Deliberate
Zoom out to OpenAI's summer pricing moves and the agentic push stops looking like a feature release and starts looking like a business model. August 10 brought premium seats for ChatGPT Business at the hundred-dollar tier. August 13 previewed Ultrafast mode for GPT-5.6 Sol, claiming up to 14x speed. September 3 cut Sol's price 20% across the API, Codex credits, and ChatGPT Work. And ads began testing in ChatGPT on August 11, all while o3 was retired and the model lineup consolidated around the work stack.
Cheaper, faster inference underneath; premium seats and ads on top. That spread only makes sense if the volume between them explodes, and the volume is agent runs. OpenAI is pricing for a world where every knowledge worker delegates constantly, which is worth internalizing: the platform's financial gravity now pulls toward more agent activity touching more sources more often. Your site is one of those sources whether you've planned for it or not.
Your Docs Are Now a Steering Layer
Funny enough, AGENTS.md files and skills now deserve a marketer's attention, which is a sentence that would have sounded absurd two years ago.
These are machine-readable instructions that tell agents how to operate: what a resource does, when to use it, what done looks like. Today they mostly live in codebases. But the pattern generalizes fast. An agent evaluating your product mid-task needs exactly that shape of information about you: what the product does, how to try it, where the API starts, what the limits are. Vendors who publish integration docs an agent can follow end-to-end become tools agents can actually use, and pages structured for agent reading become the difference between appearing in the deliverable and being a fetch error in a log nobody reads.
For what that means in practice: public docs with clean structure, quickstarts that work without a demo call, machine-parseable pricing and limits, and stable URLs an agent can cite back to the human.
What B2B Teams Should Do This Quarter
- Run the delegated-buyer test. Give an agent your buyers' actual evaluation task, in your category, and read the deliverable. Whoever's missing from it, including possibly you, just learned something a survey would never show.
- Ship agent-followable docs. Quickstart, pricing, limits, integration steps, all public and structured. This is the AGENTS.md era's version of a landing page.
- Expect and welcome agent traffic. Managed cloud agents mean more programmatic visits. Blocking them blocks the deliverable you want to appear in.
- Measure where the run ends up. The run itself is invisible, so track how engines and agents describe your brand weekly across platforms, the output side being the only side you can see. The reshuffles keep coming, and the six-engine picture diverges more every month.
Hand-running that measurement across platforms is a few hours weekly; RankControl's agents run it continuously and flag when your presence in answers moves, so the first sign of an agentic shift in your category is an alert rather than a lost quarter.
AI search traffic grew 835% this year. Is your content ready?
RankControl generates 26 content formats optimized for ChatGPT, Claude, and Perplexity. Published on your domain, matched to your brand.
The Quarter Ahead
Watch two collisions. Agents API apps meeting real workloads will produce the first hard data on what work agents actually read and cite at scale. And Astra's task-first behavior meeting OpenAI's ad tests will force the question this whole space has been avoiding: when an agent's deliverable recommends a vendor, who paid, and who verified?
The brands that thrive in that argument will be the ones whose presence was earned and machine-legible before it started. Work search is becoming agentic on OpenAI's schedule, and it isn't asking anyone's permission.
Built by the team that got cited in 48 hours.
Content generation, backlink building, AI visibility tracking, and Google rankings. One platform, zero guesswork.
