The division between conversational AI assistants and dedicated autonomous coding agents is approaching an operational end, according to Thibault Sottiaux, Head of Codex at OpenAI. Speaking in a recent interview with technology analyst Matthew Berman, Sottiaux—a former Google DeepMind research infrastructure engineer known within developer circles as “Tibo”—outlined OpenAI’s long-term technical roadmap, detailing how standalone tools like Codex will ultimately merge directly into ChatGPT to deliver a unified, highly contextualized Artificial General Intelligence (AGI) interface.
Sottiaux’s detailed accounting of OpenAI’s current internal dynamics offers a rare glimpse into the engineering philosophy of the AI sector’s dominant commercial player. The conversation covered structural bottlenecks in hardware, the operational integration of recursive self-improvement (RSI), and the cultural starkness separating OpenAI’s rapid release model from the risk-averse environment of legacy tech giants like Google.
Lessons from Google: Business Protection versus Rapid Deployment
Reflecting on his career prior to joining OpenAI in 2024, Sottiaux discussed his tenure at Google DeepMind, where he built infrastructure supporting flagship research breakthroughs such as AlphaGo. According to Sottiaux, Google had developed conversational language models—internally designated as “LM Chat”—nearly a full year before ChatGPT sparked the modern generative AI boom in late 2022.
However, Google’s institutional posture prevented these systems from reaching the public. Sottiaux noted that DeepMind was structured primarily as a fundamental research laboratory rather than a product delivery organization. Furthermore, corporate leadership’s anxiety regarding potential disruption to existing core businesses, combined with rigid centralized master plans, effectively shelved early conversational products.
“DeepMind was a very creative place, but it was not designed as an organization to deliver products,” Sottiaux explained during the interview. He emphasized that OpenAI’s competitive edge relies on tight, daily co-design between research scientists and product engineers. At OpenAI, bottom-up project creation is encouraged, prioritizing rapid public release and iterative user feedback over long-term protective holding strategies, even when doing so risks disrupting established internal revenue streams.
The Architecture of Next-Generation Agents
Addressing the current state of autonomous coding tools, Sottiaux characterized Codex in its present iteration as a transitional form. Contemporary developer workflows require users to manually curate memory stores, maintain explicit skill configuration files, and manage fragmented sub-agent networks. This constant administrative oversight constantly breaks the illusion of interacting with an autonomous partner.
The next architectural evolution of AI agents will focus on making these background mechanisms functionally invisible to the end user. Rather than requiring active file and memory management, next-generation agents are engineered to passively absorb personal workflows, long-term goals, organizational standards, and active team dynamics to proactively initiate solutions.
Furthermore, Sottiaux highlighted physical hardware limits as a primary barrier to future agent expansion. Laptops and personal workstations are fundamentally engineered around human operational boundaries—tailored for manual typing speeds, limited display multitasking, and single-user cognitive capacity. Advanced agents capable of orchestrating dozens or hundreds of concurrent background tasks will inevitably outgrow local machine environments.
As a result, agent workloads are moving decisively off local machines and into cloud environments capable of provisioning massive computing resources on demand. In this paradigm, local hardware will serve merely as an interactive portal while deep code compilation, multi-hypothesis testing, and broad environment simulation execute across distributed cloud clusters.
Ultra Fast Latency and the Convergence of Codex
Sottiaux also addressed how dramatic improvements in model throughput will fundamentally alter developer behavior. Currently, power users frequently deploy 10 to 15 parallel agent sessions to compensate for model generation latency, waiting 30 to 45 minutes for complex code runs to return completed results. This methodology imposes heavy cognitive switching costs on human operators.
With the rollout of ultra-fast inference modes—delivering token generation speeds 10 to 14 times faster than previous baseline standards—the operational dynamic shifts back toward real-time human-AI interaction. When inference latency approaches the speed of human thought, developers will no longer need to maintain complex parallel networks of sub-agents to maximize efficiency. Instead, work will return to a concentrated “flow state,” supported by instantaneous tool execution and high-speed voice dictation.
This speed dividend underpins OpenAI’s decision to integrate Codex directly into ChatGPT. Rather than maintaining separate products for writing code versus conducting general research, OpenAI views the separation as an artificial construct. Future systems will present a single, adaptive interface that dynamically reconfigures its UX, tool access, and computational depth based on the immediate context of the user’s prompt.
Recursive Self-Improvement in Practice
One of the most significant operational revelations from the interview involved OpenAI’s practical application of recursive self-improvement (RSI). Rather than treating self-improving AI as an abstract theoretical benchmark, Sottiaux revealed that OpenAI already actively deploys its most capable frontier models to optimize its underlying operational infrastructure.
Frontier models are currently tasked with refining low-level CUDA kernels, redesigning inference stacks, and tuning compute cluster distribution. This creates a compounding operational flywheel: more capable models generate higher infrastructure efficiency, which unlocks additional effective compute capacity, allowing OpenAI to train and deploy even stronger successor models.
Sottiaux concluded by framing OpenAI’s broader market strategy against competitors like Anthropic. Rather than focusing strictly on leaderboard benchmark dominance, Sottiaux emphasized that OpenAI’s primary strategic objective is maximizing distribution density—bringing frontier capabilities to the widest possible user base by eliminating usage friction and integrating developer tools into mainstream consumer platforms.