Everyone is talking about AI agents. Anthropic just showed us what is actually happening with them.
In a new research paper published this week, Anthropic analyzed millions of real-world interactions between humans and AI agents across both Claude Code and its public API. The findings cut through the hype with hard data -- and the implications for businesses deciding whether to adopt AI agents are significant.
The headline finding: AI models are already capable of far more autonomous work than people ask them to do. Anthropic calls this a "deployment overhang," and it means the bottleneck for AI agent adoption is not the technology. It is us.
Software Engineering Is Eating the Agent World
The most striking data point is how concentrated AI agent usage remains. Software engineering accounts for nearly 50 percent of all agentic tool calls through Anthropic's public API. Business intelligence, customer service, sales, finance, and e-commerce each claim only a few percentage points.
This does not mean agents are only useful for writing code. It means that software developers were the first group to build agent workflows into their daily routines, and everyone else is still figuring out where to start.
For small and medium businesses, this is an opportunity signal, not a warning. The companies that move next -- applying agents to operations, customer workflows, and internal processes -- will capture the same productivity gains that engineering teams are already seeing.
The Autonomy Numbers Tell a Surprising Story
The research tracked how long AI agents work without human intervention. The median work session in Claude Code lasts about 45 seconds -- short, focused bursts. But the outliers are where things get interesting.
Among the longest-running sessions (the 99.9th percentile), autonomous work time nearly doubled between October 2025 and January 2026, jumping from under 25 minutes to over 45 minutes. That increase was smooth across different model releases, which means it was not driven by better models alone. People are getting more comfortable letting agents run.
The trust pattern is clear. New Claude Code users fully auto-approve about 20 percent of sessions. After roughly 750 sessions, that number climbs past 40 percent. Experienced users are not reckless -- they actually interrupt the agent slightly more often (about 9 percent of work steps, compared to 5 percent for new users). They just flip from manually approving each step to letting the agent run and stepping in only when something goes wrong.
This is exactly how trust should build with any new tool, and it mirrors how businesses adopt any significant process change: cautiously at first, then with increasing confidence as results prove out.
The Agent Asks for Help More Than You Do
One of the most underappreciated findings: Claude Code pauses itself to ask clarifying questions more often than humans interrupt it. On complex tasks, the agent stops to request guidance more than twice as often as users jump in to correct course.
The top reasons Claude stops itself:
- Presenting a choice between proposed approaches (35 percent of self-pauses)
- Gathering diagnostic information or test results (21 percent)
- Clarifying vague requests (13 percent)
- Requesting missing credentials or access (12 percent)
- Seeking approval before taking consequential actions (11 percent)
Compare that with why humans interrupt:
- Providing missing context or corrections (32 percent)
- The agent was slow or excessive (17 percent)
- Already received enough help to proceed alone (7 percent)
For businesses evaluating agent safety, this data is reassuring. Modern AI agents are not runaway systems that barrel through tasks without checking. They are built to pause at decision points, especially when stakes are high. The agentic AI safety discussion is important, but the data shows that well-designed agents have meaningful built-in guardrails.
The Deployment Overhang Is Your Competitive Window
Anthropic's concept of a "deployment overhang" deserves attention. It means that today's AI models can handle substantially more autonomy than users currently grant them. According to an evaluation by METR, Claude Opus 4.5 can solve tasks with a 50 percent success rate that would take a human nearly five hours.
This gap between capability and usage creates a window for businesses willing to move faster than the average adopter. The models are not the bottleneck -- organizational readiness is.
The parallel to cloud computing adoption is useful here. In the early 2010s, cloud infrastructure was technically ready for enterprise workloads years before most enterprises moved. The companies that migrated early built lasting advantages in speed, scalability, and cost structure. The same dynamic is playing out with AI agents.
What This Means for Your Business
Anthropic's data suggests three practical takeaways:
Start with high-volume, low-risk workflows. The study confirms that agents perform best on tasks that are well-defined and reversible. Internal document processing, data summarization, scheduling, and customer inquiry triage are natural starting points -- exactly the kinds of tasks where the agentic workflow approach delivers immediate returns.
Plan for a trust curve. Your team will start by manually reviewing every agent action. That is fine. The data shows that trust builds naturally with experience, moving from step-by-step approval to autonomous operation with spot checks. Build your rollout plan around this curve rather than expecting instant full autonomy.
Act during the overhang. Nearly every industry outside software engineering is still at the starting line of agent adoption. Finance, customer service, sales, and operations each represent less than a few percent of current agent usage. That gap will not last. Businesses that establish agent-powered workflows now will have compounding advantages as adoption accelerates.
The Bottom Line
The chatbot era trained businesses to think of AI as a question-and-answer tool. Anthropic's research makes clear that the shift to agentic AI is already well underway -- but almost exclusively in software engineering.
The next wave will be agents handling multi-step business processes across every department. The models are ready. The tooling is maturing. The only question is which businesses will adopt agents during the overhang, and which will wait until the advantage has closed.
The data is in. The opportunity is now.
At BaristaLabs, we help small and medium businesses implement AI agent workflows that deliver measurable results. If you are ready to explore what agents can do for your operations, get in touch.