Key Takeaways:
Coding agents are becoming the general execution engine for knowledge work, with enterprise Codex usage in legal up 108 times since February.
Coding agents are spilling out of engineering departments into legal, sales, recruiting and marketing, with enterprise Codex weekly active users in legal up 108 times since February, according to a16z's Aug. 21 Charts of the Week report on OpenAI's enterprise signals.
"The growth slope has migrated from technical to non-technical departments," the venture firm said in the report, which tracked weekly active Codex users across OpenAI's enterprise customer base. Engineering, the original beachhead, grew just five times over the same period — a gap a16z attributed to both a high base and the spread of agents into new functions.
Sales and recruiting each grew 41 times, marketing 26 times and healthcare 24 times, the data show. Codex now accounts for about 64 percent of combined Codex and ChatGPT output tokens among enterprise customers as of June, OpenAI said, a sign compute load is shifting from answering questions to executing work.
The shift redraws the economics of enterprise software. If agents execute workflows directly against data and application programming interfaces, software value migrates from the user interface to the system of record underneath — a re-rating risk for SaaS vendors whose moat is screen complexity, and a tailwind for inference compute providers as each agent task consumes dozens of steps.
Software development offered agents three advantages: well-defined goals, a structured digital environment and easily verified results. That closed loop — observe, think, act, verify, correct — is now migrating to other knowledge work. Anthropic's study of roughly 400,000 Claude Code sessions found humans typically decide "what" while Claude decides "how," and that domain experts complete agent tasks at rates approaching software engineers. The deeper the domain knowledge, the more work each instruction completes.
The implication: scarce skills are shifting from syntax to domain knowledge, problem definition, judgment and agent orchestration. A lawyer with 20 agents can do the work of a small team; an analyst can read hundreds of earnings calls at once. Anthropic's Economic Index shows about 49 percent of occupations now have at least a quarter of their tasks covered by Claude usage, with use spreading beyond coding.
The most exposed software companies are those whose stickiness rests on complex user interfaces without proprietary data, business state or permission systems — the layers agents must call rather than bypass. Systems of record holding enterprise data and workflow state become more valuable, not less. Microsoft's 2026 Work Trend Index shows active agents in Microsoft 365 grew about 15 times in a year, and about 18 times in large enterprises, while finding that organizational factors — culture, management support, talent systems — matter more than individual tool use in producing value.
Autonomous AI is already delivering finished work. Blitzy, an enterprise coding company backed by Northzone and valued at $1.4 billion after raising $200 million, ingests legacy code and compliance rules to rebuild systems; Builders FirstSource tripled software development velocity in its first three months. XBOW, founded by GitHub Copilot creator Oege de Moor, topped HackerOne's US leaderboard last summer — the first non-human number one — and caught a firewall bypass Moderna's deputy CISO had missed. Google's Big Sleep found 20 new vulnerabilities in open-source software, and OpenAI's Aardvark caught 92 percent of known flaws in benchmark tests.
The winners may not be the model labs alone. Inference compute providers, agent runtime infrastructure, systems of record, identity and governance layers, and companies that compile industry know-how into agent workflows all sit on the new value chain. Gartner warns that by 2027, 40 percent of enterprises will demote or decommission autonomous agents because of governance gaps discovered after production incidents, so audit trails and accountability will be prerequisites, not afterthoughts. But companies tripling development speed in a single quarter are already running product on autonomous AI.
This article is for informational purposes only and does not constitute investment advice.
