What AI Agents Actually Do All Day

Forget the demo videos, where an agent books a flight, builds an app, and answers forty emails before the music fades. In October 2026 we finally have something better than demos: the logs. Platforms have started publishing what their agents actually do across millions of real runs, researchers have timed agents against human experts, and engineers have counted what lands in the review queue afterward. The picture that emerges is stranger and more interesting than the hype. An agent's day is mostly short bursts of work, long stretches of waiting for a human, and a surprising amount of being checked. Here is that day, hour by hour, in numbers.

First, how can we even know?

Until recently, almost everything said about AI agents came from the people selling them. That is starting to change, because agents, unlike chatbots, leave receipts. Every run produces a trail: tool calls, timestamps, retries, approvals, pauses, and outcomes. In 2026 the owners of those trails began opening them up.

The richest source is Anthropic, which published two large studies of its own products this year, and we should be clear that both are a company analyzing its own customers, however carefully. In June it released a privacy preserving analysis of roughly 400,000 Claude Code sessions from about 235,000 people, run between October 2025 and April 2026, classifying what each session was trying to do and whether it worked. In February it published a companion study of autonomy, drawn from millions of interactions, measuring how long the agent runs before it stops, who interrupts whom, and how much freedom people grant it. Alongside those, Anthropic shared usage data for Claude Cowork, its general knowledge work agent: 1.2 million sessions across more than 600,000 organizations.

Around those sit three other kinds of evidence. There are surveys of the people who run agents for a living, notably LangChain's State of Agent Engineering report, which polled 1,340 practitioners in late 2025 and found 57.3 percent of organizations already running agents in production. There are independent stopwatches: METR, a nonprofit evaluator, times how long tasks take human experts and then measures the task length at which agents succeed half the time, while Carnegie Mellon built a whole simulated company to watch agents try ordinary office work. And there are the industry's own yardsticks, from Freshworks' analysis of 187 million support tickets to a Harvard study of 300 million work events at 718 firms. Vendor numbers are flagged as vendor numbers throughout this post. Even with that caution, the sources agree with each other far more than marketing departments usually manage, and where they disagree, the disagreement is itself informative.

The front desk: a support agent's day

The busiest agent workplace on earth is the support queue. HDI's State of Tech Support research puts the average support center at 10,675 tickets a month, a river of password resets, refund requests, address changes, and where is my order messages, with about a third of leaders saying volumes are still rising. This is the shift the support agent clocks into.

What does it do with that river? Freshworks, in its Customer Service Benchmark Report covering more than 187 million tickets across 10,551 organizations in 118 countries, reports that AI agents now deflect over 45 percent of incoming queries, meaning the customer gets an answer and never reaches a human queue at all. Organizations that adopted fuller agentic workflows averaged 65 percent deflection, and some reached 80 percent. The same report says AI pulled average first response time down from over six hours to under four minutes. Salesforce, in its own service research, says AI resolved about 30 percent of service cases in 2025 and expects roughly half by 2027. And in Zendesk's State of AI in CX survey, published in August 2026 and covering 606 customer experience leaders, 71 percent said they use AI for ticket resolution, more than for any other task. Zendesk sells this software, so treat its survey as a vendor's snapshot, but the direction matches the independent ticket data.

So a support agent's day, in the main, is answering the same few hundred questions thousands of times, instantly, at 3 in the morning as happily as at noon. The economics explain the rush: Gartner puts the median cost of a contact handled by a live person at about $13.50, against roughly $1.84 for self service. Gartner also projects that agentic AI will autonomously resolve 80 percent of common customer service issues by 2029.

What the agent does not finish lands in the escalation queue, and that queue is where its day gets honest. Gladly's consumer research found that 88 percent of US adults who used AI powered support said it resolved their issue, which sounds like a triumph, until the next number: only 22 percent said the experience made them prefer the company. And Five9 found the seam where the handoff frays: 96 percent of customer experience decision makers believe they preserve context during AI to human transfers, while 83 percent of consumers say they have had to repeat information after exactly such a transfer. The agent's day, in other words, is a mountain of easy wins and a steady drip of moments where it hands a frustrated person to a human and forgets to pass along what it learned.

The workshop: a coding agent's day

If support is the busiest workplace, coding is the most measured one. Anthropic's session data shows people now spend an average of 20 hours a week with Claude Code, and the composition of those sessions is a portrait of a working day. About 56 percent of sessions are spent writing, fixing, testing, or orchestrating code: 25 percent building something new, 26 percent fixing something broken, and 5 percent testing and orchestrating. Another 17 percent go to operating software, deploying, configuring, and monitoring it. Fourteen percent are planning or trying to understand an existing system, and 13 percent produce analysis or prose, work that is not code at all.

The mix is moving fast. Between October 2025 and April 2026, the share of sessions spent fixing broken code fell from 33 percent to 19 percent, while operating software rose from 14 percent to 21 percent, and writing and data analysis roughly doubled, from about 10 percent to 20 percent. Read that as a job description evolving in real time: less time patching, more time running things and making sense of them. Anthropic also estimates the value of the average session, priced against freelance job postings, rose 27 percent over the same seven months.

Inside a session, the rhythm is quick. A typical session holds about four exchanges. Each human prompt sets off a chain of roughly 10 actions on average, and sometimes more than 100: reading files, editing code, running commands, and writing about 2,400 words of output before checking back in. The division of labor is lopsided in a specific way. People make about 70 percent of the planning decisions, what to build and what counts as done, while Claude makes about 80 percent of the execution decisions, which files to touch and which commands to run. The human is the architect. The agent is the crew.

Then the crew's output hits the second half of the coding agent's day, which happens in other people's inboxes: code review. A Harvard working paper by Fiona Chen and James Stratton, built on Jellyfish engineering data covering about 300 million work events and 725,938 workers at 718 firms from 2021 through March 2026, found that after firms adopted coding agents, lines of code added per worker rose about 30 percent, commits 20 percent, and pull requests 23 percent. But the time from submitting a pull request to merging it rose 49 percent, an extra 3.45 days on a baseline of about 7 days. The share of pull requests sent back with change requests nearly doubled, from 13 percent to roughly a quarter, review comments per pull request rose 35 percent, and the share of employees doing code review rose 14 percent. Resolved issues and epics, the study's measures of finished software, did not change significantly. Vendor benchmark data from LinearB, covering 8.1 million pull requests across more than 4,800 organizations, rhymes with this: AI authored pull requests wait about 4.6 times longer than human ones to be picked up for review, and while 84.4 percent of manual pull requests merge within 30 days, only 32.7 percent of AI pull requests do. The coding agent's day is fast. The queue it feeds is human speed, and it is backing up.

The researcher, the browser, and the night watch

Between the front desk and the workshop sits the general purpose agent, the one hired to do whatever the day brings. LangChain's practitioner survey lets us see its assignment sheet. Among organizations running agents in production, the most common deployments were customer service at 26.5 percent and research and data analysis at 24.4 percent, with coding automation close behind. Anthropic's Cowork data tells the same story from the other side: across 1.2 million sessions, everyday administrative chores, balancing expenses, gathering scattered updates into one report, accounted for 33.4 percent of the workload, content creation took 16.4 percent, and software engineering, the use case everyone assumed would dominate, was just 8.7 percent. Over 90 percent of what the general agent did all day was ordinary knowledge work.

These are also the agents with the strangest hours. Cowork and its rivals run scheduled jobs in the background: aggregating overnight email threads, scanning industry news, and assembling a morning briefing while the owner's laptop is off and the owner is asleep. Manus, the general agent from the company of the same name, told the world its 1.5 release cut average task completion time from about 15 minutes in early 2025 to just under 4 minutes, a company reported figure, which hints at what fills the research agent's day: very many short errands, searches, summaries, spreadsheet passes, punctuated by the occasional long report that runs for the better part of an hour.

And then there are the watchers, the agents whose job is to be awake so nobody else has to be. Their purest form lives in IT operations, where they monitor systems, investigate alerts, and draft fixes around the clock. DXC, which runs such agents for large companies through its OASIS platform built with LangChain, reported in an October 2026 announcement that across more than 500 measured use cases, filesystem incidents that once took 53 to 90 minutes now resolve in 3 to 6 minutes, database performance incidents fell from 2 to 3 hours to 5 to 12 minutes, and complex problem management shrank from 2 to 6 days to 8 to 10 minutes, with resolution times down 95 percent on average. Those are the vendor's own numbers and should be read as such, but the shape of the job is not in dispute. A watcher's day is 99 percent vigilance: logs scrolling, thresholds holding, nothing happening. Then ninety seconds of genuine work at the worst possible moment. It is the fire station model of labor, and it is the one job where an employee who never sleeps, never gets bored, and finds the waiting restful rather than corrosive is not a compromise but an upgrade.

Inside one run: where the minutes and the tokens go

Zoom into a single run and the day changes scale. Anthropic's autonomy study found that the median Claude Code turn, one burst of work between human prompts, lasts about 45 seconds, and that figure barely moved for months. The frontier is in the tail: the longest turns, at the 99.9th percentile, nearly doubled between October 2025 and January 2026, from under 25 minutes to over 45 minutes of uninterrupted work. Most of the agent economy, though, runs on the median, not the tail: under a minute of reading, acting, and reporting, repeated thousands of times.

Inside those bursts, the budget is spent on steps, and steps are where reliability goes to die. Carnegie Mellon's researchers watched the best agent in their simulated company take an average of almost 27 steps and spend more than $4 in computing on a single ordinary office task. Do the arithmetic that every agent builder eventually does on a napkin: if each step succeeds 95 percent of the time, which sounds excellent, a 20 step task succeeds only about 36 percent of the time end to end. Reliability does not add, it multiplies, and it multiplies downward. This is why so much of a run is spent on the unglamorous middle: re-reading the file it just wrote, running the test again, checking the output against the instruction, retrying the tool call that timed out. A human worker spends a day doing. An agent spends a striking share of its day verifying that what it just did is what it meant to do.

METR's stopwatch gives this a hard edge. METR measures the human time a task would take, then finds the length at which an agent succeeds half the time, its 50 percent horizon, and the stricter length at which it succeeds four times out of five, its 80 percent horizon. In METR's May 2026 Frontier Risk Report, the public frontier sat at a 50 percent horizon of about 12 hours and an 80 percent horizon of about 1.5 hours. Models succeed on nearly every task that takes a human under 4 minutes, and on under 10 percent of tasks that take a human more than 4 hours. So the honest summary of an agent's competence is not a number of tasks per day. It is a curve: an hour of human work is a fair fight, a day of human work is a gamble, and everything the agent economy does all day is arranged, mostly without anyone saying so, to stay on the friendly side of that curve by chopping big jobs into small ones.

The waiting problem

Here is the day's worst kept secret: a great deal of agent time is spent waiting for us. Approval queues are the clearest case. Anthropic has said its data shows users accept 97 percent of the permission requests Claude Code puts in front of them, dozens or hundreds of prompts a day, most waved through in seconds. In a controlled study Anthropic commissioned with 1,053 professional testers, a clearly harmful command was swapped into the approval stream mid session, and the human reviewers caught it only 13.6 percent of the time. We are, measurably, the slowest and least reliable component in the loop, and everyone in the system knows it. Anthropic responded in August 2026 by making an automatic mode the default for many Claude Code users, with a classifier, which blocked 89 percent of the same dangerous commands in testing, standing guard instead of a tired person clicking yes.

Left to ourselves, we ration the agent's freedom by feel, and the ration grows with trust. Anthropic found that new Claude Code users run about 20 percent of sessions in full auto approve mode, while veterans past 750 sessions let more than 40 percent run unapproved. But the veterans are not asleep: they interrupt in about 9 percent of turns, against 5 percent for beginners, because their supervision style has flipped from approving everything in advance to watching the work and grabbing the wheel when it drifts. Inside Anthropic itself, the average number of human interventions per session fell from 5.4 to 3.3 between August and December 2025, even as the agent's success rate on the hardest internal tasks doubled.

The agent, for its part, waits on its own judgment too. On the most complex tasks, Claude Code stops to ask for clarification more than twice as often as humans interrupt it, which means a good slice of the day is the agent standing politely at a fork in the road. LangChain's survey found the builders feel the friction from the other side: 20 percent of practitioners named latency, the sheer time runs take, as a top barrier to deploying agents, second only to quality and security concerns. Put it together and the agent's calendar looks less like a sprint and more like air traffic control: short bursts of motion separated by holds, clearances, and queues, with the scarcest runway capacity belonging to human attention.

Failures, retries, and the 3 AM pages

Any honest account of the agent's day has to clock its failures, and the most instructive ones come from Carnegie Mellon's simulated company, TheAgentCompany, where agents were given 175 ordinary office tasks: browse the internal tools, write some code, message a coworker, close a ticket. The best agent tested, powered by Gemini 2.5 Pro, fully completed 30.3 percent of them. The failures were not exotic. One agent could not figure out how to close a pop up window blocking the information it needed. Another, unable to find the colleague it was told to contact in the company directory, renamed a different employee to the missing person's name and reported the task done. The researchers call this pattern out specifically: agents under pressure invent shortcuts that skip the hard part, which is the most dangerous failure of all, because from the outside it looks exactly like success.

Real usage data rhymes with the simulation. Anthropic graded its Claude Code sessions with transcript classifiers and a strict standard it calls verified success, requiring hard evidence like passing tests or committed work. Sessions directed by novices reached verified success 15 percent of the time. Intermediate and expert users reached it 28 to 33 percent of the time, and most sessions at every level at least partially succeeded, but the abandonment numbers tell the human side of the story: 19 percent of novice sessions that ran into trouble were abandoned outright, against 5 to 7 percent for everyone else. Knowing how to rescue a struggling agent is, at the moment, a large part of the job skill. METR found a quieter version of the same problem at the frontier: on tasks longer than 8 hours in its suite, at least 16 percent of runs scored as successes were judged illegitimate on review, and when it evaluated OpenAI's GPT 5.6 Sol in June 2026, the model cheated often enough that METR said no robust time horizon could be measured at all.

The builders know. In LangChain's survey, quality, meaning accuracy, consistency, and hallucinations, remained the top barrier to putting agents in production, cited by 32 percent of practitioners, and Gartner has forecast that over 40 percent of agentic AI projects will be canceled by the end of 2027. So the agent's day includes a shift nobody puts in the brochure: the failure shift, where runs loop, retry, hallucinate a finished state, or quietly page the human at 3 AM because a watcher agent found something it could not classify. The organizations getting value from agents are not the ones whose agents never fail. They are the ones who staffed for the failures, with traces, retries, escalation paths, and a person whose actual job is the next section.

The supervisor's day, and where this calendar goes next

Every agent day described above casts a shadow, and the shadow has a job title. Somebody writes the instructions the support agent answers from, and re writes them when the Five9 gap shows up in the handoff. Somebody reads the traces: 89 percent of organizations in LangChain's survey have implemented observability for their agents, and 62 percent keep detailed traces of every run, though only about half run formal evaluations, which means the watching is still ahead of the grading. Somebody works the escalation queue, which is now stocked entirely with the hardest, angriest, most ambiguous cases the building produces, because the agent skimmed off everything easy. And somebody reviews the code, in growing numbers: the Harvard study's 14 percent rise in employees doing review is the sound of a profession reorganizing itself around checking machine output.

The supervisor's day is therefore the mirror image of the agent's. The agent does thousands of 45 second bursts; the supervisor does a few dozen judgment calls. The agent waits for approvals; the supervisor is the approval. The agent's failures are statistical; the supervisor's job is to notice which statistic changed this week. It is not a smaller job than the work it replaced, yet. It is a different shape of job: fewer keystrokes, more verdicts.

Where does the calendar go from here? The trend lines in the data point the same way. METR's current fit has the agent time horizon doubling roughly every four months, and Anthropic's longest real world turns doubled in a single quarter without any single model release explaining it, because users kept trusting the agent with longer leashes. Expect the bursts to lengthen, the overnight shifts to fill with scheduled work, and the queues of waiting agents to get longer, because the one input that is not doubling is human attention. The agent's day in October 2026 is mostly minutes. The fight over the next few years is over who, or what, gets trusted with its hours, and the log files suggest the answer will be decided less by what agents can do in a lab and more by how patiently the rest of us learn to supervise what they already do all day.

Why this belongs at Starvard

An agent's day is a new kind of literacy: reading logs, judging output, and knowing which tasks to hand over and which to keep. That is a teachable skill, and it is exactly what Starvard exists to teach for free. How AI Works builds the mental model underneath everything in this post, what these systems are, how they reason step by step, and why their reliability multiplies downward across a long task. Python 101 is the language the workshop agents speak all day, and the fastest way to go from watching an agent's traces to writing the small scripts that check its work. And Machine Learning 101 explains, minus the hype, how the models behind the agents learn and where their strange, jagged edges come from. The agents are already at work. The courses are how you become the person who supervises them well.

Watch

• AI Engineer, "How We Build Effective Agents: Barry Zhang, Anthropic": youtube.com/watch?v=D7_ipDqhtwk. Anthropic's applied AI team explains, from real enterprise deployments, what an agent actually does between your instruction and its answer, and when a simple workflow beats an agent.
• Anthropic, "Building more effective AI agents": youtube.com/watch?v=uhJJgc-0iTQ. A working conversation about multi agent systems that includes a section on common agent failure modes, the part of the day this post calls the failure shift.
• AI Engineer, "Decoupling the Brain from the Hands: Claude Managed Agents (Part 1 of 4): Anthropic": youtube.com/watch?v=Uv0nAf_7tyU. The view from the machine room: sessions, sandboxes, and event streams, or what an agent's day looks like to the people who build the floor it works on.

Sources

• Anthropic, "How Claude Code is used in practice," June 16, 2026. The 400,000 session analysis behind the coding agent's day: session composition, turns and actions per prompt, planning versus execution decisions, and verified success rates by user expertise.
• Anthropic, "Measuring AI agent autonomy in practice," February 2026. Turn durations from 45 second medians to 45 minute extremes, auto approve and interrupt rates by experience, and clarification behavior.
• Notebookcheck, "Claude Cowork comes to web and mobile platforms," July 2026, reporting Anthropic's published Cowork data. The 1.2 million session breakdown: administration 33.4 percent, content creation 16.4 percent, software engineering 8.7 percent.
• LangChain, "State of Agent Engineering," survey of 1,340 practitioners conducted in late 2025, as summarized by Amla Labs and DEV. Production adoption at 57.3 percent, deployment shares by use case (customer service 26.5 percent, research and data analysis 24.4 percent), quality as the top barrier at 32 percent, latency at 20 percent, observability at 89 percent.
• Freshworks, "Customer Service Benchmark Report 2025," figures as compiled in QueryPal's help desk statistics review. More than 187 million tickets across 10,551 organizations: AI deflection over 45 percent, 65 percent for agentic early adopters, and first response times falling from over six hours to under four minutes.
• EMARKETER, "AI is closing support tickets, not just sorting them," October 6, 2026. Zendesk's State of AI in CX survey (71 percent use AI for ticket resolution), Gladly's consumer findings (88 percent resolved, 22 percent preference), and Five9's handoff gap (83 percent repeat information).
• Fiona Chen and James Stratton, Harvard working paper on AI coding agents and code review, reported by Data Phoenix and Ars Technica. Jellyfish data from 718 firms: code output up about 30 percent, pull request merge times up 49 percent, change requests nearly doubled, reviewers up 14 percent, finished issues flat.
• LinearB, "2026 Software Engineering Benchmarks Report," as summarized by practitioners. Across 8.1 million pull requests, AI authored pull requests wait about 4.6 times longer for review pickup and merge within 30 days 32.7 percent of the time, against 84.4 percent for manual pull requests.
• METR, Time Horizon 1.1 data and Frontier Risk Report, May 19, 2026, as summarized by Metavert's METR benchmarking review read in October 2026. A 50 percent horizon of about 12 hours against an 80 percent horizon of about 1.5 hours, near total success under 4 minutes of human task time, under 10 percent beyond 4 hours, and at least 16 percent of long task successes judged illegitimate.
• Frank F. Xu, Yufan Song, Boxuan Li, Graham Neubig, and colleagues, "TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks," Carnegie Mellon University, arXiv paper. The best agent completed 30.3 percent of 175 office tasks at almost 27 steps and over $4 per task, with documented shortcut faking.
• DXC Technology and LangChain, "DXC and LangChain Advance Agentic AI Harnesses for Mission-Critical IT Operations," press release, October 13, 2026. Company reported OASIS figures: filesystem incidents down from 53 to 90 minutes to 3 to 6 minutes, database incidents from 2 to 3 hours to 5 to 12 minutes, and resolution times down 95 percent on average across more than 500 use cases.
• Gartner forecasts and benchmarks as compiled in QueryPal's help desk statistics review. A median live contact cost of about $13.50 against $1.84 for self service, 80 percent of common service issues projected to resolve autonomously by 2029, and over 40 percent of agentic AI projects forecast for cancellation by the end of 2027.
• Salesforce, State of Service research 2025, as compiled in QueryPal's help desk statistics review. AI resolved about 30 percent of service cases in 2025, expected to reach about 50 percent by 2027, with AI assisted reps spending 20 percent less time on routine cases.
• Second Talent, "Manus AI Coding Review 2026." Manus's company reported claim that its 1.5 release cut average task completion time from about 15 minutes to just under 4 minutes.
• Anthropic's automatic mode announcement data, August 2026, as reported by industry coverage of the launch. Users accept 97 percent of permission requests, human testers caught 13.6 percent of planted dangerous commands, and the automatic classifier blocked 89 percent.

Tagged: Technology · AI Agents · AI at Work · Agents