Summary
- 2026 has emerged as the year agentic engineering became standard practice inside OpenAI’s own research operations.
- OpenAI appears to be internally developing a framework around Recursive Self-Improvement (RSI), which commentators are linking to the organisation’s AGI definition.
- A notable spike in AI spend per researcher was observed internally in late July, the cause of which is not confirmed by sources.
- The normalisation of coding agents inside frontier AI labs sets a precedent that will flow through to enterprise adoption expectations.
- Security leaders should begin assessing governance frameworks for agentic AI tools before adoption outpaces policy.
Agents Are No Longer Experimental Inside OpenAI
According to commentary from researcher and technologist Simon Willison, 2026 has been the year agentic engineering genuinely took hold at OpenAI — not as a pilot or a research curiosity, but as a normal part of how the organisation’s own teams operate. Willison notes that OpenAI’s research staff are now actively using coding agents as part of their day-to-day workflow, a shift he describes as a significant inflection point. For security leaders, this matters: when frontier AI labs normalise agentic tools internally, enterprise adoption pressure tends to follow quickly.
RSI and AGI: A Terminology Shift Worth Noting
Willison also flags that OpenAI appears to be coalescing around the term RSI — Recursive Self-Improvement — as a central concept, to the point where internal communications no longer feel the need to expand the acronym. He links this to OpenAI’s evolving internal definition of AGI, and notes that both a piece from OpenAI’s research team and a new essay by Chief Scientist Jakub Pachocki, titled ‘An Alien Mind’, reference it. The significance for security executives is less about the technical definition and more about the directional signal: the organisation is openly discussing self-improving systems as a near-term operational frame, not a distant theoretical one.
An Unexplained Spend Spike in Late July
Willison highlights a chart from OpenAI’s internal data showing a marked acceleration in AI spend per researcher in late July. He offers a hypothesis — that this may correspond to when internal staff gained access to the model later released publicly as GPT-6 Astra — but is clear this is a personal inference, not a confirmed explanation. OpenAI has not provided a public account of what drove the change. Security leaders should note that unexplained capability jumps inside AI providers can have downstream implications for the tools their organisations depend on, and that version changes of this magnitude warrant re-evaluation of risk assumptions.
What This Means for Enterprise AI Governance
The picture that emerges from Willison’s analysis is one of accelerating internal adoption of agentic and highly capable AI systems within one of the world’s most influential AI organisations. For CISOs, the practical concern is not what OpenAI does internally, but the pace at which these capabilities and norms diffuse outward. When coding agents become standard inside a research organisation, they tend to become expected inside engineering teams across the industry shortly after. Governance frameworks, acceptable-use policies, and data-handling standards for agentic tools need to be in place before demand outstrips oversight.
Why it matters
The normalisation of agentic coding tools inside OpenAI’s research operation is a leading indicator of where enterprise engineering teams will be in the near term. CISOs who wait for adoption to arrive before building governance structures will find themselves reacting rather than managing. Additionally, the emergence of Recursive Self-Improvement as an operational concept — not merely a research term — signals that the capability trajectory of AI systems being integrated into enterprise toolchains may be steeper than current risk models assume. Understanding provider-side capability changes, even when they are not fully disclosed, is becoming a necessary part of third-party risk management.
What to do now
- Review your organisation’s current acceptable-use and data-handling policies to assess whether they adequately cover agentic AI tools, including coding agents used by engineering teams.
- Monitor OpenAI’s public communications and release notes for any further disclosure around the capability changes referenced in late July, as these may affect risk assumptions for OpenAI-dependent tooling.
- Begin or accelerate a governance framework specifically for agentic AI, distinct from general generative AI policy, given the different risk profile of autonomous task execution.
- Engage your engineering and research teams now to understand where agentic tools are already in use, whether formally sanctioned or not, before conducting a structured risk assessment.
