@K3v1nS0m4ny

PRIMER · AGENT AUTONOMY

AI Agents Are a Spectrum, Not a Switch

AI agents are not simply on or off. They run from low autonomy to high: fixed business workflows, browser and coding agents, and humanoid robots. Here is the spectrum, and why more freedom always needs more control.

· 16 min read Beginner
On this page
  1. Key Terms
  2. Why “Agentic” Is More Useful Than Arguing Over “Agent”
  3. Low-Autonomy Agents: Useful Because They Are Predictable
  4. Semi-Autonomous Agents: Useful Because They Can Adapt
  5. Highly Autonomous Agents: Powerful Because They Can Plan, Risky Because They Can Wander
  6. Humanoid Robots Make the Autonomy Spectrum Visible
  7. Why More Autonomy Creates More Risk
  8. Why Autonomy Levels Need to Be Designed, Not Assumed
  9. The Practical Lesson
  10. References
  11. Footnotes
In 60 seconds
  • An agent is not on or off. The useful question is not “is this a real agent?” but “how agentic is it?”, meaning how much freedom it has to decide what to do next.
  • The spectrum runs low to high. From fixed business workflows, to browser and coding and research agents, to humanoid robots, systems hold more or less autonomy.
  • More autonomy is more power and more risk. The same freedom that lets a system solve open-ended problems also makes it harder to control, test, secure, and trust.
  • Match the freedom to the stakes. Give a system the right amount of autonomy for the job, with safeguards: human approval, permissions, monitoring, and an emergency stop.

An AI agent is software that can pursue a goal through a sequence of decisions and actions, rather than simply responding once to a single prompt and stopping. A normal chatbot might answer a question about black holes, but an agent can decide that it needs background sources, search the web, open documents, compare information, draft an essay, review the draft, decide that more information is needed, gather additional sources, and then produce a more complete answer.

That distinction matters because the next wave of AI is not only about better answers. It is about AI systems that can do work across many steps. The shift is already visible in customer-support systems that resolve tickets, coding agents that modify software repositories, research agents that gather and summarize sources, office agents that manipulate spreadsheets and calendars, and physical robots that use AI models to interpret instructions and act in the real world.

The word “agent” can create confusion because it sounds binary, as if a system either is an agent or is not an agent. A better way to think about it is to ask how agentic the system is, meaning how much freedom it has to decide what to do next. Some agents follow a fixed script written by a human engineer. Others choose among tools, revise their own plans, and adapt as they go. The most ambitious systems can plan long sequences of action, operate software on a user’s behalf, coordinate with other agents, and sometimes write new code to extend what they can do.

The vocabulary

Key Terms#

A few plain definitions make the rest of this piece easier to follow. Each of these is a building block of what an agent is and what it can do.

Large language model (LLM)
The AI system that powers modern chatbots and many AI products. It is trained on vast amounts of text and other data so it can understand natural-language instructions, generate text, summarize information, write code, reason through problems, and decide which action might be useful next.
Tool use
An AI model calling outside software instead of only generating words. A tool might be a web search engine, a calculator, a code interpreter, a database, an email client, a browser, a calendar, a payment system, a PDF reader, or a robot-control module. Tool use is what lets an AI system move from answering to acting.
API
Application programming interface: a controlled doorway between pieces of software. If an agent uses a weather API, it is asking a weather service for structured weather data. If it uses a search API, it is asking a search engine for results.
Deterministic workflow
A process where the steps are fixed in advance. If the system always searches the web, always opens the top five pages, always summarizes them, and always writes a final answer, then the model may generate the content, but the route through the task was chosen by a human programmer.
Autonomy
How much decision-making freedom the system has. A low-autonomy agent may only fill in text inside a human-designed workflow. A semi-autonomous agent may choose which tool to use and when. A highly autonomous agent may decide the whole plan, revise it, and keep acting until it believes the goal has been met.
Multiagent system
A setup that uses more than one agent. One agent might research a topic, another might critique the research, another might write the article, and another might check the final answer for errors.
Teleoperation
Remote human control of a machine. This matters in humanoid robotics because many impressive robot demonstrations include some level of human control behind the scenes, even when the robot appears to be acting independently.
Embodied AI
AI placed inside a physical system, such as a robot. A chatbot only has to produce language, but an embodied AI system has to deal with cameras, sensors, motors, balance, collisions, objects, people, and messy physical environments.

A better question

Why “Agentic” Is More Useful Than Arguing Over “Agent”#

The argument over what counts as a “true agent” has never been very useful, because real systems do not fall neatly into two categories. A customer-service bot that can look up an order, issue a refund, and escalate a complaint is clearly doing more than a chatbot, but it may still be following a tightly controlled workflow. A browser agent that can navigate websites, fill forms, and compare options has more freedom, but it may still need human confirmation before making purchases or sending messages. A humanoid robot that can sort packages in a warehouse has physical autonomy, but usually only inside a narrow environment with known tasks.

Calling these systems more or less agentic is more precise than arguing over whether they deserve the label “agent.” The more useful question is how much autonomy the system has, and whether that amount is appropriate for the task, rather than whether it qualifies as a “real agent.”

That question becomes more important as agents move into higher-stakes settings. An agent that drafts a travel itinerary can make mistakes with limited consequences. An agent that sends emails, books flights, trades financial assets, changes production code, or controls a humanoid robot has a much larger blast radius. The more freedom the system has, the more important it becomes to define boundaries, permissions, audit logs, human review points, and emergency-stop mechanisms.

Predictable by design

Low-Autonomy Agents: Useful Because They Are Predictable#

A low-autonomy agent has a fixed path through a task. Imagine asking a system to write an essay about black holes. The system might first generate three search queries, then call a search engine, then open the top results, then summarize those pages, then write the essay. The model may produce fluent text, but the overall sequence was designed in advance.

This kind of agent is less glamorous than the fully autonomous systems people imagine, but it is where much of the practical business value is today. A company can build an agent that reads incoming support tickets, classifies each ticket by topic, retrieves the relevant policy, drafts a suggested answer, and sends the draft to a human representative for approval. The agent is doing useful work, but it is not deciding the whole process for itself.

Recent enterprise AI products show how valuable this level can be. Salesforce’s Agentforce,1 Microsoft Copilot Studio agents,2 ServiceNow AI agents,3 and similar tools from major software companies focus heavily on bounded business workflows: answer customer questions, retrieve account information, update records, trigger approvals, and route edge cases to people. These systems do not need to be fully autonomous to save time, because much of office work already follows repeatable patterns.

The advantage of low autonomy is control. The company knows what steps the agent is allowed to take, which systems it can access, and where a human must approve the result. The downside is limited flexibility. If the task falls outside the predefined path, the agent may fail, escalate, or produce a shallow answer.

It adapts

Semi-Autonomous Agents: Useful Because They Can Adapt#

A semi-autonomous agent can make some decisions about how to complete a task. Using the black-hole essay example again, the agent might decide whether to search general web pages, recent news, academic papers, or astronomy databases. It might decide that a summary of the basics is not enough, that it should look for recent observations from the Event Horizon Telescope, gravitational-wave research, or new papers about supermassive black holes. It might write a first draft, notice that the section on Hawking radiation is thin, search again, and improve the explanation.

This is the kind of agent people started to see more clearly when browser-using and computer-using AI systems became public. Anthropic introduced computer-use capabilities for Claude in 2024, allowing the model to interact with a computer interface in controlled ways.4 Google demonstrated Project Mariner, a browser agent built around Gemini 2.0, to show how an AI system could move through web pages and complete tasks.5 OpenAI’s Operator, and the later agent-style ChatGPT features that absorbed it, pushed the same idea into mainstream conversation, with systems that could navigate websites, fill forms, compare options, and ask for confirmation before sensitive actions.6

These products showed the difference between a chatbot and an agent in a way ordinary users could understand. If an AI tells you how to book a hotel, it is acting like an assistant who gives advice. If it opens a browser, compares hotels, checks dates, filters by budget, fills the booking form, and asks before payment, it is acting like an agent.

Semi-autonomous agents are powerful because they can adapt to what they find. They are also harder to test, because the same request may lead to different paths on different days. A website might change its layout. A search result might contain misleading information. A malicious page might include hidden instructions telling the agent to ignore the user and leak data, a problem known as prompt injection.7 A human can often ignore such tricks, but an agent that reads web pages as instructions can be manipulated if it is not designed carefully.

It plans

Highly Autonomous Agents: Powerful Because They Can Plan, Risky Because They Can Wander#

A highly autonomous agent can decide the plan, choose tools, revise the plan, and continue working over many steps. In software development, this means an agent might read a bug report, inspect the codebase, identify the likely file, write a patch, run tests, diagnose failures, change the patch, and submit a pull request. Cognition’s Devin,8 SWE-agent research projects,9 GitHub Copilot coding agents,10 and other coding-agent systems have made this style of work a major area of competition since 2024.

The same idea appears in research agents. A user can ask for a report on a technical topic, and the agent can search, open sources, compare claims, extract evidence, organize the material, draft a report, and cite the sources it used. These systems are useful because serious research is not one action. It is a chain of actions involving search, judgment, synthesis, and revision.

The hard part is that long chains create more chances for error. If an agent makes a bad assumption early, every later step may build on that mistake. If it chooses weak sources, the final report may look polished while resting on poor evidence. If it has permission to send messages, edit files, or spend money, a planning mistake can turn into a real-world consequence.

That is why highly autonomous agents need stronger controls than low-autonomy agents. They need permission boundaries, tool restrictions, human approval for irreversible actions, memory limits, logging, evaluation tests, and ways to stop execution. A powerful agent without boundaries behaves less like an employee and more like an unpredictable process that happens to have access to real tools.

Autonomy you can see

Humanoid Robots Make the Autonomy Spectrum Visible#

Humanoid robots are a useful comparison because their autonomy is visible. When software acts on a screen, the steps can feel abstract. When a humanoid robot walks, grasps, sorts, balances, or responds to voice commands, people can immediately see the difference between scripted action, remote control, and genuine autonomy.

At the low-autonomy end, many humanoid robot demonstrations rely on tightly controlled environments or teleoperation. Tesla’s Optimus demonstrations, including public events where robots interacted with guests, drew attention partly because observers questioned how much was autonomous and how much involved remote human assistance.11 That debate was useful because it reminded the public that a robot can look intelligent while still depending heavily on human operators, scripted behaviors, or carefully prepared conditions.

At the more practical low-to-middle range, humanoid robots are entering warehouses and factories in narrow roles. Figure announced a commercial agreement with BMW to test humanoid robots in automotive manufacturing, later putting its Figure 02 robot to work on a pilot line at the Spartanburg plant.12 Agility Robotics has pushed Digit into logistics and warehouse use cases, including work with major partners such as Amazon and GXO.13 These deployments are not general-purpose household robots. They are machines tested inside structured settings where the tasks, objects, floor plans, and safety rules can be constrained.

The middle of the spectrum is where robots combine perception, learned behavior, and task-specific autonomy. Boston Dynamics’ new electric Atlas, revealed after the retirement of the hydraulic Atlas, has been shown handling parts in factory-style demonstrations, using perception and motion planning rather than simply playing back a fixed dance routine.14 Figure’s robot demonstrations with speech and object handling, including the widely discussed Figure 01 video developed with OpenAI support, gave the public a glimpse of a robot connecting language, vision, and action.15 1X has shown its NEO humanoid performing household-like tasks, with EVE as its earlier wheeled model for business settings, while emphasizing how much data collection and human supervision still go into training.16

The highest-autonomy version is the household humanoid many people imagine: a robot that can enter an unfamiliar home, understand a vague instruction, locate objects it has never seen before, avoid pets and children, handle fragile items, recover from mistakes, and know when to ask for help. That is far harder than a factory or warehouse deployment, because homes are chaotic, layouts vary, objects are not standardized, lighting changes, people behave unpredictably, and the robot must be safe around vulnerable humans.

This is where the comparison becomes more useful than the old self-driving-car analogy. A car mainly moves through roads that are designed for vehicles, governed by traffic rules, and filled with signals, lanes, maps, and predictable constraints. A humanoid robot must operate in human spaces that were not designed for robots, use arms and hands near people, understand social context, and interact with thousands of ordinary objects. Autonomy in a humanoid robot is not only about navigation. It is about perception, manipulation, balance, judgment, language, and safety at the same time.

Freedom cuts both ways

Why More Autonomy Creates More Risk#

More autonomy gives an AI system more room to solve problems, but it also gives the system more room to make mistakes. A low-autonomy customer-support agent might draft a bad reply, but a human can review it before sending. A semi-autonomous browser agent might fill the wrong form, but the system can require confirmation before submission. A highly autonomous software agent with repository access might introduce a subtle security bug if it is allowed to modify and merge code without review.

The same trade-off is sharper in robotics, because physical action creates physical risk. A humanoid robot does not merely produce wrong text. It can drop an object, block a walkway, collide with a person, damage equipment, or fall. Even a small humanoid can be dangerous if it has motors, metal joints, battery systems, and enough weight to injure someone.

This is why serious robotics companies focus on constrained pilots before broad deployment. Warehouses and factories are easier than homes because the environment can be mapped, tasks can be repeated, humans can be trained to work around robots, and safety zones can be defined. A home robot must handle open-ended conditions, which means the acceptable error rate has to be extremely low.

Security risk also rises with autonomy. If a browser agent can read websites, click buttons, and access accounts, attackers can try to manipulate it through malicious pages, hidden text, fake instructions, or poisoned data. If a humanoid robot is connected to cloud services or fleet-management software, attackers may target the network, the update pipeline, the remote-control systems, or credentials. A robot fleet can multiply a single failure if many machines share the same vulnerable software.

Design the dial

Why Autonomy Levels Need to Be Designed, Not Assumed#

A strong agent does not need maximum autonomy for every task. It needs the right autonomy for the job. A refund agent might be allowed to approve refunds below a small amount but require a human for larger refunds. A research agent might be allowed to search and draft freely but not fabricate citations or use sources outside an approved set. A coding agent might be allowed to open a pull request but not merge it. A humanoid warehouse robot might be allowed to move bins in a defined area but stop when a human enters a safety zone.

This design approach is beginning to shape regulation and safety standards. The EU AI Act treats many autonomous systems as high-risk when they affect safety, employment, critical infrastructure, education, law enforcement, or access to essential services.17 The EU Machinery Regulation, scheduled to apply from 2027, updates rules for physical machines in a world where software and AI can change behavior after deployment.18 Robotics safety standards such as ISO 10218 are also being updated for modern industrial robot systems, including collaborative and more flexible automation.19

For humanoid robots, autonomy grading is likely to become as important as performance demos. A serious buyer will not only ask whether a robot can fold clothes, move boxes, or carry parts. The buyer will ask when the robot acts alone, when it asks for approval, when it stops, how it handles uncertainty, what data it records, who can take control, how updates are validated, and what happens when the network fails.

The takeaway

The Practical Lesson#

AI agents should be judged by the amount of autonomy they have, the tools they can access, the consequences of their actions, and the safeguards around them. A low-autonomy agent can still be extremely useful if the task is repetitive and the process is clear. A semi-autonomous agent can save far more time if the task requires adaptation. A highly autonomous agent can tackle complex, open-ended work, but only if its permissions, monitoring, and failure modes are designed with care.

Humanoid robots make the same lesson visible in the physical world. A robot remotely controlled by a human is not very autonomous, even if it looks impressive. A warehouse robot performing one trained task in a controlled area is more autonomous, but still narrow. A household humanoid that can safely operate around people in unpredictable spaces would be far more autonomous, and therefore far harder to build, test, regulate, and trust.

In one line

The label was never the point. What matters is how much freedom a system has, what it can do with that freedom, and whether the people who built it have matched that freedom with enough control.

Sources

References#

Product, robotics, and policy facts below are 2024–2026 snapshots from the cited primary and news sources, verified 2026-07-23.

Footnotes#

  1. Salesforce, “Salesforce Unveils Agentforce — What AI Was Meant to Be,” Sep 12, 2024. salesforce.com. ↩

  2. Microsoft, “Unlocking autonomous agent capabilities with Microsoft Copilot Studio,” Oct 21, 2024. microsoft.com. ↩

  3. ServiceNow, “ServiceNow releases its most comprehensive set of new AI innovations with the Now Platform Xanadu release,” Sep 10, 2024. servicenow.com. ↩

  4. Anthropic, “Developing a computer use model,” Oct 22, 2024. anthropic.com. ↩

  5. Google, “Introducing Gemini 2.0: our new AI model for the agentic era,” Dec 11, 2024. blog.google. ↩

  6. OpenAI, “Introducing Operator,” Jan 23, 2025. openai.com; OpenAI, “Introducing ChatGPT agent: bridging research and action,” Jul 17, 2025. openai.com. ↩

  7. OWASP, “LLM01:2025 Prompt Injection — OWASP Top 10 for LLM Applications,” 2025. genai.owasp.org. ↩

  8. Cognition, “Introducing Devin, the first AI software engineer,” Mar 12, 2024. cognition.com. ↩

  9. John Yang et al., “SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering,” May 24, 2024. arxiv.org. ↩

  10. GitHub, “GitHub Introduces Coding Agent For GitHub Copilot,” May 19, 2025. github.com. ↩

  11. TechCrunch, “Tesla Optimus bots were controlled by humans during the We, Robot event,” Oct 14, 2024. techcrunch.com. ↩

  12. Figure, “Figure Announces Commercial Agreement with BMW Manufacturing to Bring General Purpose Robots into Automotive Production,” Jan 18, 2024. prnewswire.com; BMW Group, “Humanoid Robots for BMW Group Plant Spartanburg,” 2024. bmwgroup.com. ↩

  13. Agility Robotics, “GXO Signs Industry-First Multi-Year Agreement with Agility Robotics,” Jun 27, 2024. agilityrobotics.com. ↩

  14. Boston Dynamics, “An Electric New Era for Atlas,” Apr 17, 2024. bostondynamics.com. ↩

  15. Figure, “Figure Raises $675M at $2.6B Valuation and Signs Collaboration Agreement with OpenAI,” Feb 29, 2024. prnewswire.com; New Atlas, “GPT-enhanced humanoid speaks and reasons as it works,” Mar 14, 2024. newatlas.com. ↩

  16. 1X Technologies, “NEO Home Robot,” accessed Jul 23, 2026. 1x.tech. ↩

  17. European Union, “Annex III: High-Risk AI Systems Referred to in Article 6(2), Regulation (EU) 2024/1689 (AI Act),” in force Aug 1, 2024. artificialintelligenceact.eu. ↩

  18. European Union, “Regulation (EU) 2023/1230 on machinery, applicable from 20 January 2027,” Jun 29, 2023. eur-lex.europa.eu. ↩

  19. ISO, “ISO 10218-1:2025 — Robotics — Safety requirements — Part 1: Industrial robots,” 2025. iso.org. ↩

Loading…

Sign in or create an account.

Enter your email and we will send you a sign-in link. No password needed.

or continue with

Free membership includes the newsletter — unsubscribe anytime without losing your account.