In 2026 the buzzword in AI moved from “chatbot” to agent. Instead of answering one question, an AI agent takes a goal and works on it step by step, for minutes, hours or even days. This post explains how agents work using ideas from your computer science course, then looks at two real 2026 examples, including what went wrong.
Chatbot vs agent
GCSE core| Feature | Chatbot | Agent |
|---|---|---|
| Input | One message | A goal |
| Steps | One reply | Many, chosen as it goes |
| Tools | Usually none | Search, run code, read and edit files, call other programs |
| Who decides the next step? | The user | The model, based on results so far |
The agent loop
Underneath, most agents run a simple loop, which you could write as pseudocode:
goal ← task from the user
WHILE goal not achieved AND steps < limit
plan ← model decides the next action
result ← run the chosen tool (search, code, edit)
add result to memory
IF model judges goal achieved THEN stop
ENDWHILE
report to the userThe agent loop: plan, act with a tool, observe the result, check progress, and repeat until the goal is reached or a step limit is hit. Schematic.
Notice two features you meet in A-Level computing: an iterative loop with a termination condition (including a step limit, so it cannot run forever), and state (the growing memory of what has happened so far). The big difference from a normal program is that the “plan” step is not written by a programmer: a language model chooses it at run time.
Quick check
Why does the agent loop include 'steps < limit' in its condition?
Many agents at once
A-Level extensionBig tasks can be split across many agents running in parallel, like a team. Two 2026 examples:
- Biology search: about 949 Claude agent sessions ran for 21.5 hours over public DNA databases, gathered about 200,000 enzyme gene clusters, scored 3,564 candidate families and found a candidate new enzyme system, which human scientists then studied in the lab[1]. See our post on the ART enzyme discovery.
- Mathematics: dozens of Claude agents spent 11 days writing a 13-million-line Lean proof of Fermat's Last Theorem, defining concepts and proving intermediate results that others built on[2]. See AI formalises Fermat's Last Theorem.
This is decomposition, one of the core ideas of computational thinking: break a problem into parts that can be solved separately, then combine the results.
The distinction students often miss
Powerful is not the same as reliable.
Risks to think about
- Unpredictability: the same agent can take different paths on different runs.
- Permissions: an agent that can edit files or send messages can also make mistakes with them, so sensible systems limit which tools an agent may use and ask a human before risky actions.
- Cost and energy: long agent runs use a lot of computing (the Fermat project used about six billion output tokens[2]).
- Accountability: a human still needs to be responsible for what an agent does.
Exam-style question
An AI agent is used to search a large database for patterns. Describe two ways the results could be checked before they are trusted. [4 marks]
Show model answer and mark breakdown
- repeat the search several times (or with different settings) to see if the same result is found: tests reproducibility (1 mark)
- if it is not reproducible, treat the result with caution / investigate why (1 mark)
- check the result independently, e.g. by a separate method, an experiment, or a formal checker such as a proof assistant (1 mark)
- have human experts review the result (peer review) before it is used or published (1 mark)
Examiner focus: two distinct checking methods, each described with what it shows.
Common mistake: saying 'ask the AI if it is sure'. Self-assessment by the same system is not independent checking.
Algorithms, iteration and decomposition are central to GCSE and A-Level computer science, and agents make a modern, memorable example. For how the models inside agents learn, read backpropagation explained, and for step-by-step thinking, algorithm basics.
Sources and further reading
- The Next Web: Anthropic says Claude found a new enzyme system with CRISPR-like repeats (Sep 2026) — agent sessions, hours, candidates, the ten reruns
- Anthropic: Formalizing Fermat's Last Theorem (4 Sep 2026) — dozens of agents, 11 days, tokens, caveats
- The Neuron: AI in Science and Math, September 2026 — monthly overview of AI research agents
Last reviewed: 5 October 2026. Figures are typical published values and rules can change. Exam wording varies by specification, so check yours.
Frequently asked questions
What is an AI agent?
An AI system that works towards a goal over many steps on its own: it plans, uses tools such as web search, code execution or file editing, looks at the results, and decides what to do next, repeating until the task is done or it gets stuck. A chatbot answers one message at a time; an agent keeps going.
What can AI agents actually do in 2026?
Real examples from September 2026: about 950 Claude agent sessions searched DNA databases for 21.5 hours and found a candidate new enzyme system in viruses, and dozens of Claude agents spent 11 days writing a 13-million-line computer-checked proof of Fermat's Last Theorem. In both cases humans set the goal and checked the results.
Are AI agents reliable?
Not fully. In the enzyme search, the same campaign was re-run ten times and none of the reruns found the result again. In the Fermat proof, failed attempts left behind about 7% of the non-boilerplate code. Agents can be powerful, but their output needs independent checking.
How is an agent different from a normal computer program?
A normal program follows steps a programmer wrote in advance. An agent's next step is chosen at run time by a language model, based on what it has seen so far. That makes agents flexible but less predictable, which is why testing, limits on what tools they may use, and human review matter.
Reader ratings & reviews
Rate this article across a few different things — your rating helps other readers, and helps us improve.
Was this article helpful?
Tell us what you think — a correction, a question it left unanswered, or a topic you'd like covered next.
Prefer To Talk? Book a free video call at a time that suits you: 15 minutes for a quick question, 30 for anything else. Times show in your own time zone.
About the author
Sudershan Soni
Founder & Lead Tutor at Mostak Services — an MSc-qualified Mathematics, Science, Computer Science & STEM tutor with 20+ years of professional experience, teaching students from 11+ and GCSE to A-Level and beyond, online worldwide.
Read full profile