The Signal — September 11, 2026

OpenAI spent this week doing two things that do not obviously belong together. It handed developers the machinery that runs its own long-lived agents, and its CEO told staff the company might be willing to slow down.

OpenAI put the harness behind Codex into a public API

On September 10 OpenAI released the Agents API in public beta, which puts the managed Codex harness behind a single API call. The harness is the part of an agent system that is tedious to build and easy to get wrong: it assembles context, dispatches tool calls, recovers from failures, and keeps a session coherent across hours or days of work. Until now a developer who wanted that had to write it. Now the request carries a task, a model, a set of tools and a compute environment, and OpenAI runs the loop.

Where the agent's code executes is the developer's choice. OpenAI will provision and manage a sandbox, or the agent can run on the developer's own infrastructure, or inside one of nine named partner environments: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. The built-in features are the ones that break first in homemade harnesses, and they are all aimed at duration: automatic context compaction so a task can outlive a single context window, tool search that loads definitions only when needed to keep token cost and cache intact, programmatic tool calling so the agent can run calls in parallel and filter results in code before anything re-enters context, and subagents with a configurable concurrency cap. MCP servers, custom functions and built-in tools like web search all work.

The harness itself is the open-source Codex harness, so the coordination logic is readable on GitHub even though OpenAI operates the hosted copy and versions it alongside each model release. There is no fee for the API; developers pay for tokens, tools and container time. The constraints are worth naming. MarkTechPost's writeup of the product docs notes that data residency is US-only during the beta and zero data retention is unsupported, which rules out a class of regulated workloads for now. The customer results quoted on OpenAI's announcement page were supplied by those customers. And this is a public beta, which for an agent product means the failure modes have not all been found in the wild yet.

Sources: OpenAI · MarkTechPost · Investing.com

Altman told staff OpenAI is open to slowing down, and the President said he is not worried

Bloomberg reported that at a companywide meeting this week Sam Altman told OpenAI employees the company could pace its frontier development, possibly alongside several other labs, while noting that some of those labs would not agree to it. OpenAI declined to comment on the meeting itself, but it did say on the record that it has slowed parts of model development and paused certain internal training runs recently for safety reasons. The account of the meeting rests on multiple people who would not be named, so treat the internal detail as reported rather than established. The on-record half is firmer: chief scientist Jakub Pachocki has written that AI companies should be coordinating to slow down future development as needed, and that he hopes voluntary slowdowns become commonplace until shared safety bars exist. We covered that essay on September 7. What is new is the posture moving from one scientist's published view to something the CEO floated at an all-hands, with an explicit gesture toward competitors.

Altman said in July that he had raised the need to pace development with White House officials. The White House's current answer arrived the same day as the Bloomberg story. Asked whether AI worried him, President Trump said, "No, I don't have any. I have concerns that if we don't win AI, we're going to be put in a very bad position." On the competitive gap he added, "We are leading China right now by a pretty good period. I would say a year, which is, you know, considered a lot." That one-year figure is his own characterization rather than a published intelligence or technical assessment. A coordinated slowdown needs the people who would have to bless it to believe there is room to slow.

The week's anxiety traces back to Jacob Coxon, the researcher who left Anthropic after four months and wrote that the labs are racing toward self-improving superintelligence and gambling with our lives. We covered his resignation and the reply from Anthropic's alignment science lead yesterday. What has happened since is that the argument moved inside the companies. A lab exploring whether to go slower is not a commitment, a timetable, or an agreement with anyone.

Sources: Bloomberg · The Business Times · Anadolu Agency · Hindustan Times

Anthropic modeled the economy through 2030 and put its CEO's forecast in the outlier column

Anthropic's economics team published an interactive scenario explorer on September 9, along with a technical report by Anton Korinek, Chad Jones, Szymon Sacher, Tess Cotter and Peter McCrory. We held it a day; it deserved more room than a third slot next to two other stories about the same company. The model treats every job as a bundle of tasks drawn from the Labor Department's O*NET taxonomy. AI can automate a task, help a human do it faster, leave it alone, or create a new one. Scale that across the roughly $30 trillion of value the US economy produced last year and GDP, employment and the labor share come out as results rather than assumptions.

Three scenarios are highlighted. In the modest one, AI lands about where the internet landed and is hard to pick out of macro data. In the substantial one, AI is capable of half of all knowledge work by 2030 but is not actually used for most of it; growth roughly doubles, knowledge-worker wages go flat, and workers in jobs AI does not touch see gains. The extreme scenario is the one with teeth, and the report says it would likely require recursive self-improvement plus fast adoption. Annual GDP growth reaches about 15 percent, the economy doubles every four and a half years, knowledge-worker wages fall more than 10 percent, unemployment spikes past what recessions usually produce, and labor's share of a much larger pie shrinks while capital's share climbs. Total labor income is barely changed by 2030 even though the economy is enormously bigger. The Decoder puts knowledge-worker unemployment in that run at 17.9 percent; NPR renders the same scenario as nearly 14 percent of all workers losing jobs to AI, which is a different denominator and worth keeping straight.

The report assigns no probability to any of the three. Dario Amodei warned in May 2025 that up to half of entry-level office jobs could disappear and unemployment could reach 10 to 20 percent, and those numbers sit in the extreme column. Anthropic's own model now presents its CEO's public forecast as the outlier case. The company's survey of nearly 11,000 people found the median respondent closer to the substantial scenario, with roughly 10 percent in line with the extreme one, and co-founder Jack Clark told NPR he expects the technology to keep improving fast while diffusing more slowly than people assume. The page also lists criticisms Anthropic has not resolved: the model does not track individual workers, so it says little about what displacement costs a person; it leaves out the demand effects of the data-center buildout; and more than one reviewer thought the extreme case reads as a thought experiment rather than a scenario. Anthropic agrees those are limitations. Anyone quoting the extreme run's unemployment figure should quote that agreement too.

Sources: Anthropic · NPR · The Decoder

On the Editor's Desk

We are still holding the Justice Department's inquiry into how Nvidia structured its Groq licensing deal. Axios and Reuters have now confirmed the New York Times reporting, but it is all secondhand from people who would not be named, there is no filing or docket to send a reader to, and the deal value is reported as $17 billion in some accounts and $20 billion in others. A document would fix all three problems. Nvidia's only on-record response so far is a general defense of the arrangement.

Anthropic published a threat report this week expanding its distillation allegations to seven companies, including Xiaomi and SenseTime, and roughly 190 million exchanges. We ran the NSA, CISA and FBI joint advisory on the same subject on September 9. The new counts come from Anthropic about its competitors with no independent verification attached, so we are waiting on corroboration rather than printing a number nobody outside the company can check.

Google Research's ToolGrad and NVIDIA's BioNeMo inference runtime both reported strong numbers, a 99.8 percent pass rate on tool-use data generation and a 2.90x folding throughput gain, each covered by one outlet and each measured by the vendor. Google's Mantis vulnerability toolkit stays held for a different reason: Google documents it as a demonstration, and the story will be what it finds when someone outside Google points it at real code.