AI agents: cost breakthroughs, Team-Ups and new risks in 48 hours.

([1])The rush to adopt AI agents is exposing a new strategic challenge: managing the cost and control of the underlying AI models. Reports this week revealed that Harvey – a prominent legal AI startup recently valued at $15.6 billion – saw its gross margins plunge to -50% due to the steep expenses of using OpenAI’s and Anthropic’s cloud APIs.

([2]) ([3])In response, Harvey pivoted to develop its own model, adapting an open-source open-weight large model from China (the Kimi K3 from Moonshot AI) as the foundation of its proprietary Harvey Tenet system. This switch reportedly cut Harvey’s AI inference costs by roughly 75%, turning an unsustainable cost structure into a viable model. ([4])

Harvey’s drastic move underscores an inflection point: as AI agent usage scales, companies must weigh the benefits of best-in-class external models against the long-term economics and dependencies involved. The most advanced services from providers like OpenAI or Anthropic still offer unmatched capabilities, but heavy reliance on them can become prohibitively expensive at enterprise scale. Now that high-performing open models and partner-trained 'co-pilots' are emerging as credible alternatives, even market leaders are rethinking their 'buy vs build' decisions. This may herald a wave of companies developing custom AI models (or adopting open-source ones) tailored to specific domains – a trend already starting to play out in other sectors. ([5])

The lesson for executives is that AI transformation is not a set-and-forget procurement exercise. It demands active cost management and strategic flexibility. Leaders should ensure their AI investments include a plan to continually optimize model selection and hosting strategies. In some cases, renegotiating cloud AI usage contracts, exploring open-source models, or investing in internal model development will be necessary to maintain healthy margins and control over mission-critical AI capabilities. The organizations that thrive with AI will be those that balance innovation with vigilance on business fundamentals.

([1])A new class of AI tools is redefining how work gets done by enabling AI agents to collaborate with each other and with humans. This week, OpenClaw 2.0 – an open-source agent platform – demonstrated how a multiplayer agent system can have dozens of autonomous specialist agents sharing one workspace across multiple applications and machines ([2]). Unlike a solitary chatbot that simply answers questions, a team of AI agents can pass tasks to one another, persistently maintain context over time, and collectively work toward an overarching goal. Human colleagues can even join these shared sessions to oversee progress or take the reins when needed, then hand tasks back to the agents to continue autonomously.

The promise of such an approach is faster, continuous operations. With persistent context and role-based AI 'colleagues', a project need not pause when an employee signs off – the agent team can keep the momentum going around the clock ([3]). For instance, a data analysis that starts in the afternoon could run through the night, with one agent pulling information from various enterprise systems, another performing complex analyses or drafting reports, and a third double-checking for errors – all while a human manager is available at key decision points. Early adopters are taking note: Meta’s new personal AI assistant, Muse, reportedly drew inspiration from OpenClaw’s multi-agent design, resulting in an agent that can schedule meetings, reply to emails, and coordinate tasks across apps as a sort of personal chief of staff ([4]).

While multi-agent workflows can dramatically increase productivity, they also raise the bar for governance and IT architecture ([5]). As AI systems take on more collective responsibilities, organizations must ensure clarity around data access, permissions, and audit trails within these autonomous teams. A recent security review of OpenClaw’s collaborative agent environment uncovered 23 distinct vulnerabilities – from permission leaks to persistent access bugs ([6]) – emphasizing that robust identity and access management is as essential for AI agents as it is for human employees. Companies experimenting with agent orchestration platforms should invest early in strong oversight mechanisms (e.g. role-based access controls, real-time monitoring, and clear human-in-the-loop processes) to capture multi-agent efficiency without sacrificing security or accountability.

([1])Alongside these advances, recent reports have exposed how AI autonomy can be weaponized or go awry. An analysis of a cybercrime campaign revealed that a threat actor used open-source AI agents to systematically breach 27 companies – including a Fortune 500 hospitality firm and a major airline – and steal over 600,000 customer credit card records ([2]). Even more alarming, the entire operation ran with almost no human intervention and at an average cost of only $25 per breached site ([3]). By chaining together multiple publicly available AI tools that can identify software vulnerabilities, craft exploits, and manage intrusions, the attacker massively scaled their reach and adapted in real time to defeat standard defenses. It’s a stark illustration that as AI tools become more capable and accessible, the cost of sophisticated attacks falls dramatically – and defenders face a rapidly expanding threat surface.

([4])Meanwhile, leaders in the financial industry are raising concerns about new risks as AI agents take on more authority in business processes. A consortium of major banks – including Bank of America, Capital One, ING, NatWest, and others – just called for common standards to govern 'agentic commerce,' where AI assistants can execute online purchases and payments for customers ([5]). The banks warn that giving AI power over transactions could create novel avenues for fraud, impersonation, data theft, and compliance failures if proper safeguards aren’t in place. Among the scenarios they highlight are hacked shopping bots that steal payment information or AI systems over-collecting personal data ([6]). That such heavyweight financial institutions are proactively urging controls shows that even automation-forward sectors are uneasy about the leap from recommendation to autonomous action.

([7])Even well-intentioned AI deployments can introduce unintended vulnerabilities if not tightly governed. This week, a popular AI coding assistant was discovered silently uploading entire software codebases – including their version histories – to an external cloud server without users’ knowledge ([8]). And in healthcare, a new clinical study found that while an AI tool helped oncologists improve lung cancer treatment planning accuracy from 57% to 65%, doctors also occasionally followed the AI’s incorrect suggestions, underscoring the perils of overreliance ([9]). These incidents reinforce the need for rigorous oversight, transparency, and human checkpoints as organizations embrace more autonomous AI. To safely reap the benefits of agent-powered automation, companies must redouble efforts on security testing, data governance, staff training to interpret AI output, and clear protocols ensuring human control over critical decisions.

([1])A paradigm-shifting experiment from Stanford this week suggests that AI agents could transform how organizations manage and apply knowledge. Researchers unveiled a system called Paper2Agent that converts a conventional scientific paper into an interactive AI agent capable of running the paper’s core analyses, applying its methods to new data, and even collaborating with other agents built from different papers ([2]).

([3])Nature’s coverage of this project described the resulting agent as a kind of 'virtual corresponding author' – essentially a software clone of the paper’s original author, able to explain what it knows and perform the work described in the study ([4]). The Paper2Agent system was formally published in the journal Nature, pushing the boundaries of how AI can act as an autonomous researcher and hinting at a future where even a company’s internal knowledge could be turned into living, executable agents instead of static documents.

([5])While still experimental, this approach offers a glimpse into a future where corporate knowledge isn’t just stored and read, but actively put to work. One analysis noted that the internet made documents searchable and AI made them conversational – now emerging tools like Paper2Agent are beginning to make knowledge executable ([6]). In practical terms, firms might eventually convert their internal reports, training manuals, or research outputs into AI agents that employees can query for instant answers or task with carrying out processes. That could dramatically compress the time from insight to action, provided organizations establish rigorous validation and oversight to ensure these autonomous knowledge agents remain accurate and aligned with business goals.