What is agentic AI? Autonomy as a dial, not a label
Agentic AI means systems that pursue goals over many steps with little supervision. The four dimensions of agenticness and the math of compounding errors.

Agentic AI is AI that pursues a goal over many steps with limited human supervision: it plans, calls tools, reads the results and adjusts, instead of answering once and stopping. OpenAI’s governance paper from December 2023 defines agenticness as “the degree to which a system can adaptably achieve complex goals in complex environments with limited direct supervision”, which makes it a matter of degree. A chatbot with one search step is a little agentic; a coding agent that works for an hour on its own is very agentic.
Where did the term agentic AI come from?
The adjective spread through 2024. Andrew Ng’s letter in The Batch on March 20, 2024 opened a series called “Agentic Design Patterns” and named four of them: reflection, tool use, planning and multi-agent collaboration. The number that made people pay attention was on the HumanEval coding benchmark: GPT-3.5 zero-shot scored 48.1%, GPT-4 zero-shot 67.0%, and GPT-3.5 wrapped in an agent loop reached “up to 95.1%”. The loop mattered more than the model upgrade.
In December 2024 Anthropic’s Building effective agents used “agentic systems” as the umbrella term for both fixed workflows and model-directed agents. That’s the usage I follow: “agentic” describes the family, and an AI agent is the most autonomous member of it.
What makes a system more or less agentic?
The OpenAI paper, Practices for Governing Agentic AI Systems (Shavit, Agarwal, Brundage and others), argues there is no clean line between “agents” and other AI systems, and breaks agenticness into four dimensions:
- Goal complexity: how hard the goal would be for a person, and how wide a range of goals the system can take on.
- Environmental complexity: whether the work is cross-domain, multi-stakeholder, long-horizon or spread across several external tools.
- Adaptability: how well the system reacts to novel or unexpected circumstances.
- Independent execution: how reliably it reaches its goals with limited human intervention.
I find this more useful than any yes-or-no definition, because it maps onto design decisions. A support bot that routes tickets to one of five queues is low on all four. A deep research agent that searches, reads dozens of pages and writes a report is high on environmental complexity and independent execution. When someone says their product is “agentic”, I ask which of the four dials they turned up, and what they did about the risk that came with it.
There is also a measurement of how far the independent-execution dial has moved. METR’s paper Measuring AI Ability to Complete Long Software Tasks (March 2025) measures the length of task, in human time, that a model completes at 50% success. It put Claude 3.7 Sonnet at around 50 minutes and found that this horizon had doubled roughly every seven months since 2019. Longer horizons are what turned agents from demos into products in 2025 and 2026.
How is agentic AI different from generative AI?
Generative AI produces content: text, code, an image. Agentic AI uses a generative model as the decision-maker inside a loop that takes actions. The model is often the same; what changes is that its output is fed to tools, and the tool results are fed back to the model.
The practical difference shows up in three places:
- State. A generation is stateless. An agentic run carries history across steps, so you have to manage what’s in the window, the discipline I write about in what context engineering is.
- Side effects. A bad generation is a bad paragraph. A bad action is a deleted table or a sent email.
- Variance. A generation costs roughly the same every time. An agentic run can take two steps or forty.
Why do errors compound in agentic systems?
Anthropic puts the trade-off in one sentence: “Agentic systems often trade latency and cost for better task performance.” The cost side is two pieces of arithmetic, and it helps to see them with numbers. This script makes two assumptions, stated in its comments: each step has the same chance of silently going wrong, and each step adds a fixed number of tokens to a history that is resent every turn.
# agentic_math.py: why autonomy costs more than it looks. Python 3, stdlib only.
# Assumptions (mine, for illustration): every step either succeeds or silently goes wrong
# with the same probability, and nothing catches the error. Each step adds 1,500 tokens
# (tool call + result) to a history that gets resent, on top of a 3,000-token system prompt.
STEPS = [1, 5, 10, 20, 50]
print("P(run has no bad step), per-step reliability p")
print("p " + "".join(f"{str(n) + ' steps':>11}" for n in STEPS))
for p in (0.99, 0.95, 0.90):
print(f"{p:<6.2f}" + "".join(f"{p ** n:>11.0%}" for n in STEPS))
print("\nInput tokens billed for the whole run (history resent every step)")
for n in STEPS:
billed = sum(3_000 + 1_500 * k for k in range(n))
print(f"{n:>3} steps {billed:>9,} tokens {billed / (n * 3_000):>5.1f}x a one-shot call per step")
Output (Python 3.14.5):
P(run has no bad step), per-step reliability p
p 1 steps 5 steps 10 steps 20 steps 50 steps
0.99 99% 95% 90% 82% 61%
0.95 95% 77% 60% 36% 8%
0.90 90% 59% 35% 12% 1%
Input tokens billed for the whole run (history resent every step)
1 steps 3,000 tokens 1.0x a one-shot call per step
5 steps 30,000 tokens 2.0x a one-shot call per step
10 steps 97,500 tokens 3.2x a one-shot call per step
20 steps 345,000 tokens 5.8x a one-shot call per step
50 steps 1,987,500 tokens 13.2x a one-shot call per step
The first table is why a model that is right 95% of the time per step finishes a clean 20-step run only about a third of the time, if nothing checks its work. The second is why a 50-step run costs about 13 times more per step than a single call: the history keeps growing and you pay for it again on every turn. Prompt caching cuts the price of the repeated prefix but doesn’t change the shape.
Real agents beat the first table because they aren’t blind. Each tool result is a check: a failing test, an empty query result, a 404. The ReAct results and Anthropic’s post both point at grounding the model in what the environment returns at every step, and that’s what turns p ** n into something that can recover. The design job in agentic AI is mostly adding cheap checks between steps.
What does agentic AI look like in production?
By late 2026, the clearest cases are the ones with fast, objective feedback:
- Coding agents, where a compiler and a test suite grade every step.
- Research agents, which search, read and cite, with the citations as a partial check.
- Operations agents in support or IT, usually with a person approving anything irreversible.
The pattern is consistent: autonomy goes up where verification is cheap and stays down where mistakes are expensive. The OpenAI paper’s practices follow from the same logic: constrain the action space, require approval for high-stakes actions, make agent activity legible, monitor automatically, and keep the system interruptible.
What I would do
Treat agenticness as a budget you spend per dimension, and spend it only where you have a check. Concretely:
- Start with the least agentic design that works, usually a workflow with fixed steps, and add model-directed steps only where the path can’t be predicted.
- For every step you give the model, name the signal that tells you it went wrong: a test, a schema check with structured output, a second model as a judge, or a person.
- Put a hard cap on steps and tokens per run, and alert on runs that hit it.
- Keep irreversible actions behind human-in-the-loop approval until your logs show the agent gets them right.
For a search product like rSearch (DeepSeek R1 for reasoning, Serper for retrieval) the useful question is the same: at which steps can the system check itself, and at which does it have to trust the model?
FAQ
What is agentic AI in simple terms?
AI that works toward a goal by itself over several steps, choosing actions, using tools and reacting to what happens, instead of giving one answer to one prompt.
Is agentic AI the same as an AI agent?
Close but not identical. An AI agent is a system where the model directs its own loop. “Agentic” is a property that comes in degrees, so a fixed workflow with one model-chosen step is agentic without being a full agent.
What are the four agentic design patterns?
In Andrew Ng’s March 2024 series: reflection (the model critiques its own output), tool use, planning (the model writes and follows a multi-step plan) and multi-agent collaboration.
What are the risks of agentic AI?
Compounding errors over many steps, unpredictable cost and latency, actions with real side effects, and prompt injection through tool results. Most mitigations limit what the agent can do or add a check between steps.
Sources
- Shavit, Agarwal, Brundage et al., OpenAI, Practices for Governing Agentic AI Systems, December 2023
- Andrew Ng, The Batch, Agentic Design Patterns Part 1, March 20, 2024
- Anthropic, Building effective agents, December 19, 2024
- Kwa, West, Becker, Deng et al., METR, Measuring AI Ability to Complete Long Software Tasks, arXiv, March 18, 2025