Box CEO Aaron Levie: Code Cost Reduction Increases Demand for Engineers
Box CEO Aaron Levie stated that the use of programming agents will far exceed previous expectations. As code costs decrease, software will become more useful, and the leverage of engineers will significantly increase, resulting in a need for more engineers, not fewer.
He listed practical scenarios including: creating entirely new categories of tools, writing systems for companies that previously could not afford software teams, using agents to protect against network risks, automating life sciences research and development, handling massive data in complex workflows, upgrading legacy systems and infrastructure, and "hundreds" of other use cases. The core judgment is that as code becomes cheaper, the market will choose to create more software rather than stop hiring coders.
Levie previously conducted a thought experiment with the same logic: ten years ago, life sciences companies cut laboratory automation and large-scale data processing projects due to fixed team costs; when each engineer's output becomes 2x or 5x, most companies he interacted with chose to expand hiring rather than downsize. He mentioned at O'Reilly's AI Codecon that just because agents can write code does not mean the field becomes simpler; engineering will become more technical, as it is necessary to determine which parts to assign to large models and which must be locked into testable, deterministic code.
Box's own path is tied to this judgment. The company provides content platforms to about two-thirds of the Fortune 500 and will launch Box Automate in 2026, allowing agents to handle processes such as invoice extraction and document data extraction, built into enterprise packages with additional sales for higher tiers. Levie told Reuters that there are already dozens of projects internally that "would not have been initiated without AI." Box is also transforming its API into a file system for agents and integrating multiple generations of models in AI Studio for complex enterprise task assessments.
He distinguished between two types of diffusion speeds: the economic value of programming work directly corresponds to digital output, and task scale can be extended in a single session, thus agent programming growth is nearly vertical; sales, legal, and medical tasks must first change workflows, otherwise agents cannot run continuously. He named programming agents such as Claude Code, Devin, Codex, Factory, Cursor, and Replit, which already have sandboxes, can write and run code, connect directly to APIs/CLIs, and have long-term memory, while knowledge work agents are hindered by permissions, identity, and contextual fragmentation.
In market mechanisms, buyers are non-tech industries that previously lacked software budgets and enterprises needing to transform legacy systems; sellers are programming agents, cloud sandboxes, enterprise content platforms, and system integrators. Funding is shifting from "selling SaaS by seat" to "agent consumption + governance," with engineering roles transitioning from writing code to orchestrating, accepting, and integrating agents into business systems. Beneficiaries are platforms that can hand over unstructured enterprise data to agents, as well as toolchains like Cursor; those under pressure are traditional suites that rely solely on seat fees and are not friendly to agents. Event-driven changes come from leaps in model capabilities rather than single order announcements.
There is currently no unified exchange-like transaction data on the enterprise side, with observation points focusing on agent token consumption, enterprise content platform usage, and the conversion of additional sales after companies like Box integrate agent functionalities into existing packages.
Source: Public Information
ABAB AI Insight
Levie dropped out of USC in 2005 to found Box, surviving the first round of SaaS consolidation through enterprise file synchronization, turning the company into a content layer for the Fortune 500. In the face of generative AI, he did not transform Box into a chat window but rewrote the file repository into a work memory callable by agents: permissions, auditing, and unstructured document retrieval take precedence over conversation. He publicly opposes the notion that "agents will replace SaaS," advocating for the decentralization of deterministic business processes and non-deterministic agents, due to the emergence of agent leaks and erroneous operations in production environments. This aligns with his path of transforming Box from a storage provider to a workflow platform in the 2020s: first occupying data positions, then selling automation.
The capital path is to compress R&D into executable enterprise content for agents rather than building large models. Box Automate integrates into existing enterprise packages, charging incrementally; engineering budgets are being used to revive projects previously rejected due to costs. The motivation is that the seat-based model becomes ineffective when "the number of agents reaches 100 to 1000 times the number of people," necessitating a shift to usage and governance-based fees. Strategically, whoever controls enterprise files and permission maps can turn models like Claude, GPT, and Grok into auditable process workers, rather than allowing clients to migrate their data out.
Analogous examples include Salesforce turning CRM into an application platform, ServiceNow transforming work orders into a process operating system, and Palantir making data into actionable entities. Cursor, Devin, and Factory are expanding engineer-side capacity, while Box is supplying context on the enterprise side. The industry is transitioning from tool proliferation to control layer competition: programming agents have already exploded, and the next step is determining who provides agents with identity, permissions, and verifiable outputs. Life sciences, legal, and finance are repeatedly highlighted because these industries have low software density, data is in documents, and they previously could not afford complete engineering teams.
Structural judgments belong to the technological substitution that triggers a reconstruction of the industrial chain. The mechanism is not that "coding jobs disappear," but that as the marginal cost of code decreases, the software demand curve shifts outward, with the bottleneck changing from "can it be written" to "is there context, permission, and acceptance criteria." Therefore, pricing power shifts from selling development hours per head to charging based on callable data assets and governance capabilities; demand for engineers rises with the number of software projects that can be initiated, while enterprises that have not completed digitalization are forced to first purchase migration and integration services rather than laying off staff.
ABAB News · Cognitive Laws
- The first reaction to cost reduction is to do more, not hire less.
- Work that can be verified is automated first; work that cannot be verified is more expensive.
- Agents do not lack computing power; they lack callable enterprise memory.