The agent company has a shape now: small team, harness, and a kill list

The bottleneck moved

Box chief executive Aaron Levie argues that the founder's job has changed shape. He says a team of 3 to 10 people using AI agents can now produce the output of a company of 50 to 200 people, and that the bottleneck for founders has shifted from doing the work themselves to managing and reviewing the agents that do it. Levie calls the current startup environment the most different from previous eras that he has seen in 27 years. m:…0435

Levie expects hundreds of new companies worth $1 billion to $10 billion or more to be built on automating specific job functions with AI agents, and he says AI progress is currently more valuable for enterprise use cases than for consumer ones. He also argues that large incumbents will struggle to adopt agents because their existing processes are too rigid, and that AI does not fundamentally change the distribution model for startups. m:…0435

His prescription for builders is specific: the best way to build an agent startup is to have a founder with deep, direct experience in the exact job function being automated. That claim is worth holding against what companies actually report doing internally, because the inside of an agent deployment looks less like a single clever model and more like plumbing. m:…0435

Five pieces every harness needs

LangChain's account of its own internal agents describes five primitives that a company harness has to get right. The harness must integrate with existing communication platforms like Slack. It must settle identity: agents act either on behalf of a user or under fixed service-account credentials. It must use both global and user-specific memory tiers. It must run code in sandboxed environments connected to secret proxies. And it must handle the fact that a large percentage of agent runs are scheduled or triggered by background events rather than started by a person. m:…2244

The examples make the primitives concrete. LangChain says its GTM Agent monitors emails, responds to leads and sends daily briefs; OpenSWE drafts and reviews pull requests; and Content Studio drafts marketing assets including blogs, images and videos. Agents can break out into their own Slack threads so they do not pollute human conversations, and LangChain says its Managed Deep Agents product provides infrastructure for channels, crons, sandboxes, memory and identity, with LangSmith used for debugging decisions, evaluating changes and one-click deployment. m:…2244

Identity turns out to be the decision that shapes the rest. OpenSWE acts on behalf of the user when messaged, but uses its own identity for background code reviews, while Content Studio uses fixed credentials for Twitter, restricted to a whitelisted set of users. The pattern is that one harness serves several identity models at once, chosen per task rather than per company. m:…2244

The same pattern at Revolut

Revolut chief executive Nik Storonsky describes a version of the same operating model at much larger scale. He says Revolut has automated 90% of its customer support team using AI, and that its engineering speed has increased by 30% because of AI tools. The company is also developing a foundational AI model called Pragma, trained on transactional data. Storonsky says he aims to build the number one bank in at least 100 countries, and that fintech companies without a banking license are relegated to niche audiences. m:…6942

Storonsky's management method echoes the harness logic: Revolut uses a staging process where bets are funded, tracked and sunsetted if they fall into the bottom 25% of performance. He also argues that professional chief executives in banking are inherently risk-averse and prioritize job protection over innovation, which is his explanation for why founder-led companies move faster. His own venture fund, Quantum Light, uses machine learning models to select investments, so the automation story extends to capital allocation itself. m:…6942

The portfolio evidence

The venture portfolio news points the same way, at smaller scale. Social Leverage reports that Sunlight AI raised $17.4 million in seed funding to help independent medical practices manage their business operations, and that Elio Mortgage raised $5.1 million after rebranding from Boost Mortgage. Fiscal.ai, another portfolio company, launched Wall St. Estimates and Agentic Estimates, with an API and MCP that provide estimates on nearly 10,000 stocks. The firm also says it is integrating AI throughout its own tech stack. m:…8943

This suggests the agent-era company is converging on a recognizable shape: a small team, a founder who knows the job being automated, a harness that handles identity, memory, sandboxes and scheduling, and a ruthless process for killing bets that fall behind. The open question is whether that shape works as well outside coding and support, where the tasks are less structured and the failure modes are harder to see. Levie's own caveat is relevant here: incumbents struggle not because they lack models, but because their processes are rigid, and a harness is, in the end, a process. m:…0435 m:…2244