Safer agents, OpenAI expands its network and an alignment alarm
AI agents improve drastically in real work environments, OpenAI bets $150M on partners and a new startup says alignment is going badly.
Today’s most important AI news
Today the spotlight is shared between concrete evidence of progress and warning signs. On one side, an updated benchmark shows that AI agents for work environments have improved radically in two years — with specific data that changes the conversation about how ready they are for production. On the other, safety researchers launch a startup warning that AI alignment is “not on track”, the very same day OpenAI announces a partner network with $150 million behind it. A day of contrasts worth reading in full.
Agents at work: from 43% to 89% in two years, and with fewer dangerous errors
In March 2024, the best agent available for work tasks — GPT-4 on the WorkBench benchmark — completed 43% of assigned tasks and took unwanted actions (like sending an email to the wrong person) in 26% of cases. Those numbers made it hard to justify autonomous deployments at any serious company.
A new paper revisiting that same benchmark in June 2026 tells a different story: Claude Opus 4.8 now completes 89% of tasks and makes harmful errors in only 2.5% of cases. It’s not an incremental improvement — it’s a jump in category. The study also highlights something that breaks a frequent argument in the safety debate: capability and safety advanced together, not in conflict.
For product managers and automation teams, this has direct implications. The threshold of 89% completion with 2.5% serious errors is starting to be comparable to human performance on many routine administrative tasks. The question is no longer whether agents can work in work environments, but which specific processes justify the remaining human supervision and which don’t. The next step is understanding what kinds of tasks make up the 11% that still fails.
OpenAI launches its Partner Network with $150M to accelerate enterprise adoption
OpenAI today announced the OpenAI Partner Network, a program with a $150 million investment aimed at helping global partners accelerate AI adoption in companies. The proposition is clear: OpenAI wants to build an ecosystem of integrators, consultancies and solution providers that bring its models inside organizations that don’t have the technical capacity to do it on their own.
The move replicates a well-known pattern in the enterprise software industry — the same one Salesforce used with its AppExchange or SAP with its network of SIs (System Integrators). OpenAI is clearly betting that the next growth phase won’t come from adding individual users but from penetrating corporate accounts through trusted third parties. For companies in Latin America, this means there will soon be more local options for implementing solutions based on OpenAI’s models without depending on internal technical teams.
The timing isn’t coincidental. Microsoft, Google and Anthropic have been building their own implementation networks for months. OpenAI formalizes its own at a time when competition for corporate digital-transformation budgets is getting more intense. For developers and tech consultancies in the region, it’s worth reviewing what being part of this network means and what kind of certifications or requirements it involves.
”Alignment is not on track”: researchers launch Sequent
A group of researchers linked to the UK AI Security Institute launched Sequent, a new AI safety startup, with a direct premise: current alignment work isn’t keeping up with the pace of model development. The phrase that sums up their stance, published in Import AI, is blunt — “alignment is not on track” — and it comes from people with technical credibility, not activists outside the industry.
What sets Sequent apart from other players in the safety space is its focus on “under-resourced research bets” — areas they consider critical but that the mainstream of the big labs isn’t prioritizing enough. It’s an implicit critique of the model where safety is funded by the same lab that has commercial incentives to move fast.
For executives and product leaders making adoption decisions, this kind of signal deserves attention even if it doesn’t cause panic. The argument isn’t that current models are dangerous today, but that governance infrastructure and verification methods aren’t scaling at the same pace as capabilities. Put another way: 89% of tasks completed on WorkBench is good news, but knowing why the remaining 2.5% fails — and guaranteeing it doesn’t fail in unpredictable ways — is still a problem not fully solved.
In a nutshell
What connects today’s three stories is a tension that defines this moment in the industry: performance numbers are improving so fast that the “agents aren’t ready” narrative is becoming less and less tenable, but the mechanisms for understanding and controlling those systems aren’t advancing at the same pace. OpenAI builds a network to bring AI to more companies; agents show they can work with near-human reliability; and at the same time, serious researchers warn that we’re building on safety foundations nobody has finished verifying. These aren’t contradictory stories — they’re three faces of the same problem any organization adopting AI over the next 12 months will have to navigate.
Sources used: WorkBench Revisited: Workplace Agents Two Years On (https://arxiv.org/abs/2606.13715), Introducing the OpenAI Partner Network (https://openai.com/index/introducing-openai-partner-network), Import AI 461: “Alignment is not on track” (https://jack-clark.net/2026/06/15/import-ai-461-alignment-is-not-on-track-frontiercode-and-synthetic-research-interns/)