In a recent episode of O’Reilly’s This Week in AI, Andreas Welsch, founder and chief human AI officer at Intelligence Briefing, and Matt Palmer, head of developer experience at Conductor and LinkedIn Learning educator, discussed the practical implications of US export restrictions on frontier models and architectural alternatives teams can pursue.
When the API disappears
The conversation began with the immediate fallout from the US government’s limits: Anthropic’s Fable 5 and Mythos Preview were restricted to 100 handpicked US organizations, and OpenAI applied similar constraints to GPT‑5.6, capping early access at roughly 20 organizations. For many practitioners, those models effectively vanished from their stacks.
Andreas argued — echoing concerns common among European tech leaders — that while the restrictions may stem from policy issues, the core problem is infrastructure. If your stack depends on a single frontier model that can become unavailable without notice, you’ve introduced a hard dependency into your architecture rather than a simple vendor relationship.
From a builder’s perspective, Matt noted that those who used Fable 5 before the restrictions could feel the capability gap versus the next available option. That gap becomes a business risk if competitors retain access you don’t.
O’Reilly’s guests framed the lesson as architectural: teams should design for multivendor flexibility the way they design for database portability or cloud independence. Anthropic has said it expects access restrictions to evolve rapidly, but planning as if they will resolve quickly is itself a risky bet.
The delegation trap
The episode also examined the rise of agentic development and a related rise in cognitive fatigue. Developers delegating more work to coding agents report higher exhaustion: more agents mean more sessions to monitor, more approvals to give, and more half‑finished work to review.
Andreas and Matt referenced stories of engineers checking in on agents around the clock — at children’s soccer games or in bed. The promise of “it runs while you sleep” often turns into managing multiple concurrent workflows, akin to supervising a mid‑sized team. Matt described how many developers now act as managers of agents, constantly sending messages and checking progress, and that coding — once a focused, relaxing activity — feels more fragmented.
Andreas cited a Harvard Business Review study of a 200‑person software company showing that as AI tools became more capable, role boundaries blurred: product managers did prototyping, developers handled design. The tools expanded what was possible, which then expanded what teams treated as necessary — increasing workload rather than decreasing it.
He also pointed out that traditional corporate delegation is a learned skill, supported by frameworks (what is the task, the goal, data to use, characteristics of good output, time budget). Many professionals building with AI today are improvising delegation protocols without training. The industry’s investment in tooling has outpaced investment in the organizational skills required to use those tools well. More capable agents do not automatically reduce load; they redistribute it in ways that are harder to see and manage. Long‑term success will favor practitioners who can set scope clearly, check output efficiently, and protect focused work time.
One API call, many models: Sakana Fugu
The episode’s technical focus was Matt’s walkthrough of Sakana AI’s Fugu, a coordinator model/multi‑agent system from the Tokyo‑based lab. Fugu accepts a query via a single OpenAI‑compatible endpoint, routes the request to a pool of frontier models, assembles a team of specialists, and returns a synthesized result — all with orchestration hidden behind one API call.
Matt explained the architecture: a lightweight coordinator model assigns roles (one model plans the approach, another implements, a third verifies). The process can be recursive, with sub‑tasks reentering the same orchestration at smaller scale. Sakana calls this learned orchestration, and they reference two papers — “TRINITY: An Evolved LLM Coordinator” and “Learning to Orchestrate Agents in Natural Language with the Conductor” — that explore learned routing and coordination rather than hand‑designed workflows.
He also demonstrated quick integration paths: Fugu acts as a drop‑in OpenAI‑compatible endpoint via curl, through the Codex harness with a one‑line installer, and via the open‑source OpenCode harness on OpenRouter.
Sakana claims its orchestration extracts better performance from existing models: Fugu’s Ultra model scores comparably to Fable 5 on agentic benchmarks such as Terminal‑Bench, and its pricing is identical to GPT‑5.5. Whether those performance claims hold across broader real‑world workloads will be determined by the community, but the portability argument remains: because Fugu can orchestrate models from multiple providers, a restriction on any single model won’t necessarily take the system down, and teams can opt specific providers out.
Sakana launched Fugu ten days after the US export restrictions on Fable 5 and Mythos took effect, explicitly pitching AI sovereignty. That design choice directly addresses the initial problem Andreas raised: for teams in regions with access uncertainty (Europe is currently limited pending regulatory compliance, for example), multi‑provider orchestration provides resilience.
A hardware parallel: Qualcomm and Modular
The week also saw Qualcomm’s announced acquisition of Modular for roughly $3.9 billion. Modular’s platform enables AI models to run across different chip architectures — NVIDIA, AMD, and custom ASICs — without requiring developers to rewrite code per architecture. The deal reinforces a broader market trend: portability and architecture‑agnostic layers are becoming strategic investments across the entire stack.
What’s next
Guests on the show urged organizations to prioritize multivendor portability and to develop the organizational skills needed to delegate effectively to agents. The next episode of This Week in AI airs Monday, July 6, from 10:00–10:30am EST, when Christina Stathopoulos will review the latest AI developments. Episodes are available live via the O’Reilly learning platform (new users can try a 10‑day free trial) and on YouTube, Spotify, Apple, and other podcast platforms.
