The AI Brief

OpenAI published "Research acceleration: The view inside OpenAI" on September 6, disclosing that by mid-August 2026 its research organization used 3.1 agent-workdays of effort for every one human workday, a ratio derived from spend, throughput, and task classification. The company declared it had reached the goal Sam Altman announced in October 2025: fielding an automated research intern capable of carrying out well-defined tasks under human direction, including multi-day assignments, by September 2026. OpenAI's next declared milestone is a full automated AI researcher by March 2028.

By mid-August, the median researcher ranked by agent usage was running more than $600 a day of inference at API prices. OpenAI applied a task taxonomy published by Epoch AI, breaking development into six phases: Decide, Design, Build, Run, Analyze, and Communicate. The dominant use case remained research and infrastructure code, alongside technical troubleshooting and monitoring runs, while high-level planning continued to account for a minimal fraction of agent output tokens. Coding agents excelled at troubleshooting internal research infrastructure; multiple teams that previously held office hours to help researchers troubleshoot experiments noted declining attendance in 2026, and one stopped holding sessions entirely.

OpenAI reports that agents still require significant human steering, especially as task complexity rises, and that in the last six months more than half of successful 4-to-8-hour tasks involved one or more human interventions. OpenAI frames the piece explicitly as a transparency commitment around recursive self-improvement and calls on other labs to publish similar internal data. No independent audit of the methodology has been announced. The disclosure arrived three days after GPT-6 Astra shipped, a sequencing that observers including Simon Willison noted may reflect internal access to a pre-release Astra driving the sharp acceleration in agent spend visible in late-July internal charts.

Snowflake posted product revenue of $1.49 billion, up 37% year over year, in results reported September 2, 2026. CEO Sridhar Ramaswamy told Bloomberg Technology that its CoCo coding agent drove half the Q2 beat. Management raised full-year FY27 product revenue guidance to $6.07 billion, or 36% growth, from a prior $5.84 billion at 31% growth, the second consecutive raise from the initial $5.66 billion set in February 2026.

CoCo now has 9,100 accounts, an increase of over 2,000 during the quarter. Cortex AI sits inside more than 9,100 accounts, and every agentic workload built on top drives more governed queries into the platform underneath. Snowflake is not primarily priced per user; its model is consumption-based, with customers paying for compute, storage, data transfer, and AI usage, and AI work is enormously compute-hungry, running on the data customers have already placed there. The company recorded a net loss of $191.7 million, or 55 cents per share, smaller than the net loss of $297.9 million, or 89 cents per share, one year ago.

Caveat Revenue attribution between CoCo and the core platform is management's own decomposition; no independent breakdown is publicly auditable. The split is worth holding onto because it says the AI contribution is real, but not the whole thing.

The headline from OpenAI's September 6 disclosure, 3.1 agent-workdays per human workday, is qualified by structural constraints the company embeds in the same document. Agents still require significant human steering, especially as task complexity rises, and in the last six months more than half of successful 4-to-8-hour tasks involved one or more human interventions. High-level planning continued to account for a minimal fraction of agent output tokens , meaning the strategic layer of research, what to investigate and why, remains a human function.

Coding agents excel at troubleshooting internal research infrastructure, which addresses one meaningful bottleneck to research progress. That is a real gain, but a narrow one: infrastructure debugging is a bounded, well-specified task class. The broader claim, that agents are now conducting research, is not what the data shows. OpenAI is explicit that humans establish research priorities, judge results, and decide whether to scale or pause a line of work. The post frames the disclosure as a transparency commitment around recursive self-improvement and calls on other labs to publish similar data , an ask that implicitly acknowledges the field lacks any external reference point against which to check OpenAI's own measurements.

The practical implication for operators deploying agents on knowledge work: the productivity gains are real and concentrated in execution tasks with well-defined success criteria. The gap between agent capability and autonomous research judgment has not closed, it has simply become more precisely measurable.

Managed Agents appear to be the headline OpenAI is building toward for DevDay 2026, according to a report published September 7, 2026 by TestingCatalog, which found evidence in the company's own codebase for creating agents, configuring environments, and enabling skills and plugins. The agents interface is not live for anyone yet; what exists are strings and structures in shipped code, a reliable signal of intent and an unreliable signal of timing.

The codebase suggests functionality very similar to Anthropic's offering, including creating different agents and environments, managing them, and enabling specific skills and plugins. The new platform feature is expected to let users build, configure, and deploy AI agents with customizable environments, skills, and plugins. TestingCatalog has a track record on OpenAI pre-announcements, including a correct call on Agent Builder before DevDay 2025.

DevDay 2026 runs on September 29, 2026, at Fort Mason in San Francisco; the opening keynote begins at 10 a.m. Pacific and features Sam Altman. New for 2026, OpenAI follows the San Francisco flagship with DevDay Exchanges in eight cities, Bengaluru, Tokyo, Seoul, Paris, Berlin, London, São Paulo, and Mexico City. Model updates and security announcements are also expected alongside Managed Agents.

Anthropic has decided against acquiring AI startup Decart, people familiar with the matter said. Anthropic had been exploring a deal and performed due diligence on Decart, but ultimately walked away. The companies may still pursue other opportunities to collaborate. Decart had been valued at approximately $6 billion in the potential deal, which would have brought DOS chip-optimization technology inside Anthropic's inference division. First covered in Vol. I, No. 103 (August 13).

Separately, Anthropic is evaluating whether to build its own internal billing, payment, and fraud-detection infrastructure. Job listings indicate the AI startup is seeking engineers and managers to evaluate the development of proprietary billing and treasury tools. The company's Billing Platform team is running build-vs-buy evaluations across its financial infrastructure. The open role's scope includes ownership of pricing primitives, payments infrastructure, contracts, entitlements, revenue pipelines, and reconciliation, with the job description specifically referencing evaluating build-vs-buy decisions while "leaning heavily on third-party providers such as Stripe."

The move highlights potential risks for Stripe, which relies heavily on booming AI firms for growth. Beyond consumer payments, Anthropic is looking to consolidate its banking relationships and build custom production-grade treasury applications. The build signal aligns with Anthropic's pre-IPO S-1 preparation, where reliance on a single third-party payment processor for token-based billing at $65 billion annualized revenue would represent a disclosed vendor-concentration risk.

Disclosure: Claude, which generates this brief, is built by Anthropic.