anthropic's ai agent research: what the data really says - By Sourav Mishra (@souravvmishra)
a simple breakdown of anthropic's 2026 research on claude code. most agents still need humans.
anthropic just dropped some research on how people use claude code.
it turns out that 73% of tool calls still have a human checking them. and only 0.8% of actions are permanent. most of this is used for software engineering.
in this post, i, sourav mishra, will break down these numbers and explain what they mean for building ai apps.
the headline numbers
-
human oversight: about 73% of the time, a human is watching or approving what the agent does. people aren't just letting bots run wild. they are staying in the loop.
-
irreversible actions: only 0.8% of actions are permanent. most tasks can be undone or fixed. this means you only need extra security checks for a tiny fraction of what the bot does.
-
domains: about half of all agent use is for coding. the rest is for security, finance, and research. these are serious jobs, not just fun demos. you need to build with clear limits and audit logs.
-
session length: people are using agents for way longer now. the longest sessions doubled in just a few months. agents are doing real, multi-step work. we need to design apps that handle long tasks without crashing or overspending.
what this means for builders
if humans are checking 73% of actions, your app needs a great review screen.
show the user exactly what the bot is about to do before it happens. use diffs, confirm buttons, and undo features.
don't add friction everywhere. just add it to that 0.8% of permanent actions.
since sessions are getting longer, you must use timeouts and step limits. in the vercel ai sdk, i always use stopWhen: stepCountIs(N). this stops the bot from running up a massive bill.
you can see how i do this in my agentic chatbot guide. users want safe boundaries, so give them exactly that.
aligning with real usage
the data shows that people want to oversee their agents.
so, make it easy for them. require a click to approve anything destructive. let them see the bot's plan before it acts. this totally lines up with the safety rules in my ai agent security post.
if you are building for the agentic society, the rule is simple: use limits, keep tools safe, and keep humans in the loop for big stuff.
key takeaways
- most agents need humans: 73% of actions are checked by a person. only 0.8% are permanent.
- serious work: half of agent use is for coding. the rest is finance and security.
- longer sessions: people use bots for longer tasks now. always use step limits and timeouts.
- build safely: make it easy to review actions. require human approval for permanent changes. see my security post for more.
written by sourav mishra, full stack engineer for next.js and ai.
frequently asked questions
q: do people just let ai agents run on their own? not really. anthropic found that 73% of tool calls involve human oversight. most usage is closely watched.
q: are agent actions permanent? only about 0.8% are. the vast majority of tasks can easily be undone or fixed.
q: what are people using agents for? about 50% of it is for software engineering. the rest is for finance, security, and research.
q: how should i design my app based on this? use step limits for long tasks and make it super easy for users to review what the bot is doing before it acts. read my agentic chatbot guide for code examples.