Autonomous AI agents, explained for people who have to work with them

Agent, assistant or workflow: three words that get mixed up

An assistant answers inside a conversation you steer, one turn at a time. A workflow runs a fixed sequence of steps that a person designed in advance, with a model filling in some of them. An agent is given a goal, decides which steps to take, calls tools (search, files, a browser, an API) and keeps going until it judges the goal met or hits a limit.

Anthropic's engineering guide draws the line between "workflows", where models and tools "are orchestrated through predefined code paths", and "agents", where models "dynamically direct their own processes and tool usage".^[1] C The same guide recommends "finding the simplest solution possible, and only increasing complexity when needed", which is unusual advice from a company selling the complex option, and worth keeping in mind for every section below.

The risks differ. An assistant can be wrong; you read the answer and decide. A workflow can be wrong in a predictable place; you test that step. An agent can be wrong in a place nobody designed, because it chose the path itself.

The vocabulary is also moving. US searches for "autonomous ai agents" average about 6,600 a month and grew roughly 220% a year over three years, yet fell 65% year on year as attention shifted to "agentic ai", while "what is an ai agents" grew 511% to 22,200 searches a month (DataForSEO, retrieved 2026-11-09). People are asking the definitional question in larger numbers than ever. Our page on AI agents versus agentic AI covers the naming shift; this one covers what the systems do.

Five levels of autonomy

There is no standard scale for agent autonomy, so this is our framing, built for people who have to decide how much to hand over. Each level adds one thing the system does without asking.

Level	What the system does on its own	Example	Where review belongs
0. Assistant	Answers; you act	A chat window drafting an email you then send	Every output
1. Suggests actions	Proposes concrete steps or tool calls; you approve each	A coding assistant proposing a diff	Every action
2. Executes bounded steps	Runs a single tool call or short chain after one approval	"Find the three transcripts and summarise them"	Before the first action and at the end
3. Multi-step with checkpoints	Plans and runs a chain, pausing at defined gates (send, pay, delete)	A research agent that gathers, drafts and stops before publishing	At the gates and at the end
4. Unsupervised within a budget	Runs to completion within token, time and permission limits, then reports	A nightly agent that triages a shared inbox into drafts	Sampled, after the fact
5. Self-directed	Sets its own sub-goals and acquires new tools or permissions	Not a product you can buy today in a form we would recommend	Not recommended for work output

Most products sold as agents in 2026 operate at levels 2 and 3. The practical question is not "is this an agent" but "which level am I buying, and where are the gates".

What agents complete unsupervised: the evidence table

The claim worth testing is simple: given a realistic task and no human in the loop, how often does the agent finish it correctly? Three independent benchmarks and four workplace studies answer that, with grades.

Setting	What was measured	Result	Grade	Source
τ-bench (tool-use agents in retail and airline customer service, with policy rules)	Task success and consistency across repeated runs	Even state-of-the-art function-calling agents "succeed on <50% of the tasks" and are inconsistent, with pass^8 below 25% in retail	B	[2]
GAIA (general assistant questions requiring browsing, tools and reasoning)	Exact-answer accuracy	Humans 92% versus 15% for GPT-4 with plugins, on questions "conceptually simple for humans"	B	[3]
OSWorld (369 real computer tasks across operating systems and apps)	Task success in a live desktop environment	Humans above 72%, best model 12.24%, "primarily struggling with GUI grounding and operational knowledge"	B	[4]
Customer-support assistant, 5,172 agents at one firm	Issues resolved per hour with AI suggestions (level 1, human sends)	About 14–15% more on average; 30% or more for the least experienced, near zero for the most experienced	A	[5]
Consultants, 758 participants, GPT-4 as assistant	Speed and correctness inside versus outside the model's competence	12.2% more tasks and 25.1% faster inside the frontier; 19 percentage points less likely to be correct outside it	B	[6]
Experienced developers, 246 real tasks, early-2025 agentic coding tools allowed	Time to complete	19% slower with AI, while believing they were 20% faster	B	[7]
106 human–AI experiments, meta-analysis	Human plus AI versus the better of either alone	Combinations performed worse on average (g ≈ −0.23), with losses in decision tasks and gains in content creation	A	[8]

Three things to take from the table. The benchmark numbers are from the papers' publication dates (2023–2024) and newer models score higher; we have not verified current leaderboard figures, so treat the gap, not the exact percentages, as the finding. Second, the strongest evidence of benefit comes from level-1 use, where a person still sends the output. Third, the only randomised study of unsupervised-style agentic tools on real work (the developer study) found a slowdown, not a speed-up, in a sample that is small and unusually expert.^[7] None of this says agents do not work. It says the completion rate on open tasks is far below the completion rate implied by a demo.

Where they fail, and why the failures are predictable

Four failure modes account for most incidents we have read about, and three of them are older than large language models.

Inconsistency. The same agent on the same task does not produce the same result every time. τ-bench's pass^k metric exists to measure this, and the authors report that reliability over eight attempts falls below a quarter in retail.^[2] B A demo shows you one run.

Grounding. Agents that act on screens or in unfamiliar software fail at the interface, not the reasoning: clicking the wrong element, misreading a state, not knowing an application's conventions.^[4] B This is why agents inside one vendor's suite tend to work better than agents driving arbitrary software.

The jagged frontier. Tasks that look similar to a human sit on different sides of the model's competence, and the person using it cannot tell which side they are on.^[6] B An agent that handled yesterday's contract summary may mishandle today's because of one clause type it has never seen.

Automation bias in the reviewer. This is the failure mode on the human side. In a flight-simulation experiment, people given an automated aid made more errors when the aid was wrong than people with no aid, even with contradicting information on screen.^[9] B A systematic review of 74 clinical studies found users following incorrect automated advice at non-trivial rates, with trust, task complexity and time pressure as the mediators.^[10] A An integrative review in Human Factors frames complacency and bias as attentional effects of trust and workload that fall with accountability and training.^[11] A A 2025 survey of 319 knowledge workers found higher confidence in generative AI associated with lower self-reported critical-thinking effort; that is self-report, so we grade it C, but the direction matches the older literature.^[12] C Our interactive explainer, Your brain on AI, walks through what that research does and does not show.

The cost model: tokens, tool calls and human review

Vendor pricing pages show the price of a model. The cost of an agent has three parts, and the third is usually the largest.

1. Model cost. Tokens in and out, multiplied by the number of steps. An agent that reasons for ten steps and re-reads its context each time can consume ten to a hundred times the tokens of a single assistant answer for the same task.
2. Tool and platform cost. Each search, API call or automation run is billed by someone. Automation platforms meter runs or tasks; the counts matter more than the per-task price once an agent loops.
3. Human review. Minutes spent briefing, checking and correcting. If the completion rate is 50%, half the runs need correction or a re-run, and the review time scales with the error rate, not the licence fee.

A rough but honest way to estimate: expected cost per completed task = (model + tool cost per run) ÷ completion rate + review minutes × your hourly cost. With a 50% completion rate the first term doubles. With review skipped, the error cost lands downstream instead. The inputs vary by an order of magnitude between tasks, so we put no currency figure here; the OS Builder asks for yours and shows the result.

How to read a vendor claim

Agent marketing follows a pattern. These six questions separate a capability from a demo.

• Which level? Does the system act after each approval, at gates, or to completion? If the page does not say, assume level 2 and test from there.
• Completion rate on your tasks, not theirs. Ask for pass^k on a sample of your own work, or run ten tasks yourself and count. A single success rate without a repeat count hides the inconsistency problem.^[2]
• What happens on failure? Does it stop and report, retry silently, or improvise? Silent retries are where budgets disappear.
• Where are the permissions? Read, draft, send, pay, delete. A system that cannot be run with draft-only rights has no safe first month.
• What leaves your environment? Which data goes to which model provider, and under what training and retention terms.
• Who has measured it independently? A vendor case study with "hours saved" is a claim, not evidence. Benchmarks with published methods and peer-reviewed field studies are evidence, and the ones we have are in the table above.

Our comparison of the best AI agents for work applies these questions to named products, and AI coding agents covers the one domain where the evidence is thickest and most contradictory.

What changes by 2028 under scenarios A, B and C

The adoption data show intent well ahead of results. In Microsoft's 2025 Work Trend Index, 82% of 31,000 surveyed leaders said they were confident they would use "digital labour" to expand capacity within 12 to 18 months.^[13] A McKinsey's November 2025 survey, read through secondary coverage, puts 23% of organisations at the stage of scaling an agentic system and 39% experimenting.^[14] B Gartner predicts that over 40% of agentic-AI projects are cancelled by the end of 2027 for cost, unclear value or weak risk controls, and separately that 33% of enterprise applications include agentic AI by 2028; both are analyst estimates read from press-release headlines and we grade them C.^[15] C

In our Future Demand Radar, "AI agents and delegated work" scores 59.7 out of 100 in the baseline scenario (p10–p90 band 55.2–64.1) and is classed "build now". Under the conservative scenario A it scores 56.0 and stays "build now"; under the accelerated scenario C it reaches 61.7. Our baseline scenario expects search demand for the category to reach about 1.7 times today's level by 2028 (band 1.5–2.0) and about 3 times by 2030; scenario A expects 1.3 times by 2028 and scenario C 2.2 times. Forecasts are model outputs with stated uncertainty, not facts.

What the scenarios mean for a person deciding today:

• Scenario A (conservative). Agents stay features inside suites; level-2 and level-3 systems in one vendor's environment are the norm. Skills to build: briefing and acceptance testing.
• Scenario B (baseline). Delegated agents become routine in knowledge work by 2028–2030, with a 2027–28 trough if the cancellations Gartner predicts arrive. Skills to build: the same, plus run logs and cost control.
• Scenario C (accelerated). Capable agents absorb much routine knowledge work; the human job moves to judgement, orchestration and verification. Skills to build: knowing your own frontier and designing checks for work you no longer do yourself.

In every scenario the review skills pay off; only the amount of delegated work differs.

Glossary

• Agent. A model-driven system that chooses its own steps and tool calls toward a goal.
• Agentic AI. Marketing and research umbrella term for agents and systems of agents.
• Tool call (function call). A structured request from the model to run an external action and return the result.
• Orchestration. Code or a platform that sequences models, tools and gates; the "workflow" layer.
• pass^k. The probability that an agent succeeds on all k repeated attempts at a task; a measure of reliability, not just ability.^[2]
• Grounding. Correctly mapping the model's intent onto the real interface or data in front of it.
• Automation bias. The human tendency to accept an automated output over contradicting evidence.^[9]
• Guardrail / gate. A rule that stops the agent for human approval before specified actions.

What to do next

If you work with agents, your job is to pick the level, place the gates and measure the completion rate on your own tasks. The OS Builder has an AI-delegation section that asks about your tasks, privacy constraints and experience, suggests which task types to try first at which level, and estimates the review time you should budget. Capability moves quarterly; the Productivity Signal briefing tracks the benchmark and adoption numbers in the table above and tells you when they change.

Frequently asked questions

What is the difference between an AI agent and an AI assistant?

An assistant answers inside a conversation you steer turn by turn; you act on the result. An agent is given a goal, decides which steps to take, calls tools such as search, files or an API, and keeps going until it judges the goal met or hits a limit. Anthropic draws the same line between workflows on "predefined code paths" and agents that "dynamically direct their own processes".^[1] C

What can AI agents actually do unsupervised today?

Less than demos suggest. On τ-bench, state-of-the-art tool-using agents succeeded on fewer than half of customer-service tasks and were inconsistent across repeated runs.^[2] B On OSWorld, the best model completed 12.24% of real computer tasks against more than 72% for humans at publication.^[4] B Newer models score higher; the gap is the finding, not the exact figures. Bounded, checkable tasks with a human send step are where the evidence of benefit is strongest.

What are the main ways AI agents fail?

Four patterns: inconsistency (the same task succeeds on one run and fails on the next), grounding errors in unfamiliar interfaces, the jagged frontier (tasks that look alike sit on different sides of the model's competence), and automation bias in the person reviewing the output, who accepts wrong results more readily the more often the agent has been right.^{[6] [9] [10]} A

How much does running an AI agent cost?

Three parts: model tokens multiplied by the number of steps and retries, tool and platform fees per run, and the human minutes spent briefing and checking. With completion rates below 50% on open tasks, the expected cost per completed task roughly doubles and review becomes the largest line. Vendors quote only the first part.

Are the "five autonomy levels" an industry standard?

No. There is no agreed scale for agent autonomy; the levels on this page are our framing, built so that a buyer can ask which level a product operates at and where its human gates are. The useful question is not whether something is an agent but how much it does between approvals.

Will autonomous agents replace knowledge work by 2028?

Our baseline scenario expects demand for agent-related work to reach about 1.7 times today's level by 2028 and delegated agents to be routine in knowledge work by 2028–2030, with the category scoring 59.7 out of 100 on our radar (band 55.2–64.1). Gartner, by contrast, predicts over 40% of agentic-AI projects are cancelled by end-2027, an analyst estimate we grade C.^[15] C Forecasts are model outputs with stated uncertainty, not facts.

Sources and evidence

1. Building effective agents — Anthropic (2024), vendor page C www.anthropic.com/research/building-effective-agents (retrieved 2026-11-09)
2. τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains — Yao, Shunyu; Shinn, Noah; Razavi, Pedram; Narasimhan, Karthik (2024), report B doi.org/10.48550/arXiv.2406.12045 (retrieved 2026-11-09)
3. GAIA: a benchmark for General AI Assistants — Mialon, Grégoire; Fourrier, Clémentine; Swift, Craig; Wolf, Thomas; LeCun, Yann; Scialom, Thomas (2023), report B doi.org/10.48550/arXiv.2311.12983 (retrieved 2026-11-09)
4. OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments — Xie, Tianbao; Zhang, Danyang; Chen, Jixuan; Li, Xiaochuan; Zhao, Siheng; Cao, Ruisheng; et al. (2024), report B doi.org/10.48550/arXiv.2404.07972 (retrieved 2026-11-09)
5. Generative AI at Work — Brynjolfsson, Erik; Li, Danielle; Raymond, Lindsey (2025), observational A doi.org/10.1093/qje/qjae044 (retrieved 2026-11-09)
6. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua, Fabrizio; McFowland III, Edward; Mollick, Ethan R.; Lifshitz-Assaf, Hila; Kellogg, Katherine; Rajendran, Saran; Krayer, Lisa; Candelon, François (2023), RCT B doi.org/10.2139/ssrn.4573321 (retrieved 2026-11-09)
7. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — Becker, Joel; Rush, Nate; Barnes, Elizabeth; Rein, David (METR) (2025), RCT B doi.org/10.48550/arXiv.2507.09089 (retrieved 2026-11-09)
8. When combinations of humans and AI are useful: A systematic review and meta-analysis — Vaccaro, Michelle; Almaatouq, Abdullah; Malone, Thomas (2024), meta-analysis A doi.org/10.1038/s41562-024-02024-1 (retrieved 2026-11-09)
9. Does automation bias decision-making? — Skitka, Linda J.; Mosier, Kathleen L.; Burdick, Mark (1999), RCT B doi.org/10.1006/ijhc.1999.0252 (retrieved 2026-11-09)
10. Automation bias: a systematic review of frequency, effect mediators, and mitigators — Goddard, Kate; Roudsari, Abdul; Wyatt, Jeremy C. (2012), systematic review A doi.org/10.1136/amiajnl-2011-000089 (retrieved 2026-11-09)
11. Complacency and Bias in Human Use of Automation: An Attentional Integration — Parasuraman, Raja; Manzey, Dietrich H. (2010), systematic review A doi.org/10.1177/0018720810376055 (retrieved 2026-11-09)
12. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — Lee, Hao-Ping (Hank); Sarkar, Advait; Tankelevitch, Lev; Drosos, Ian; Rintel, Sean; Banks, Richard; Wilson, Nicholas (2025), observational C doi.org/10.1145/3706598.3713778 (retrieved 2026-11-09)
13. 2025 Work Trend Index Annual Report: The year the Frontier Firm is born — Microsoft WorkLab (2025), report B www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-fro (retrieved 2026-11-09)
14. The state of AI in 2025: Agents, innovation, and transformation (as reported by ppc.land) — McKinsey & Company (QuantumBlack) (2025), report B ppc.land/most-companies-still-pilot-ai-programs-despite-widespread-ado (retrieved 2026-11-09)
15. Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 (press release; analyst estimate, headline verified via search only, page returned 403) — Gartner (2025), report C www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts (retrieved 2026-11-09)