AI Agent
Software that pursues a goal by taking its own steps — deciding, acting, and reacting — instead of answering once and stopping.
Reviewed July 10, 2026Stable
When not to use it
- When a single well-crafted prompt does the job. Agents add planning loops, tool calls, retries, and failure modes. If one call answers the question, one call is the architecture.
- For anything irreversible without a human in the loop. Sending emails, moving money, deleting records — an agent that's right 95% of the time is a system that's wrong every twentieth action, unsupervised.
- When you can't afford non-determinism. The same input can take a different path each run. If your users or auditors need reproducibility, an agent is the wrong shape.
Reach for something else instead
- A workflow — fixed steps, model calls at specific points. Boring, debuggable, and correct for the majority of "agentic" projects.
- A single prompt with tools when you need one lookup, not a plan.
- Human-in-the-loop review for anything consequential. Slower on paper, faster once you count the incidents.
This entry is part of a longer guide: What is AGI?
Read more on the blog
- Why AI agents fail: the seven failure modesGartner predicts over 40% of agentic AI projects will be canceled by 2027. The failures follow patterns, seven of them. The taxonomy: what breaks, why, which real incident proved it, and which control would have prevented it.
- How AI agents actually work: the loop behind the hypeEveryone is building "agents" and almost nobody explains what makes one. Not the marketing, the actual machinery: the loop, the four components, and the single property that separates an agent from a chatbot with a fancy system prompt.
- When not to use an agentGPT-3.5 inside a structured workflow scored 95.1% on a coding benchmark. GPT-4 running free scored 67%. The structure was worth more than two generations of model improvement, and most tasks called agentic do not need an agent.
- What an agent costs, and where the money actually goesToken prices fell 67% in a year and enterprise AI bills rose anyway. The reason is a multiplication nobody models: a ten-turn agent session costs roughly fifty times a single call, not ten.
Further reading
- Yao et al. (2022), ReAct: Synergizing Reasoning and Acting in Language Models — the interleaved reason-then-act loop most agent frameworks are built on.
- Schick et al. (2023), Toolformer — models learning when to call a tool, rather than being told.
- Shinn et al. (2023), Reflexion — self-critique loops, and an honest look at where they stop helping.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Giving an agent more autonomy than the task requires, then being surprised by the blast radius.
- No budget or step limit. Loops that can't terminate are the classic agent failure — and the bill arrives regardless.
- Treating tool errors as edge cases. In production, tools fail constantly; how the agent handles a failed call is the product.
At a glance
FieldAI Agents
Core loopobserve, plan, act
Key riskcompounding error over many steps
Needstools, memory, guardrails
DifficultyIntermediate → Advanced
Often compared with
Agent vs. Workflow — decides its own steps vs. follows fixed, predefined steps.
Where this sits
24 concepts come first. Understanding it opens up 18 more.
11Levelsteps in
24Needs firstconcepts
18Opens up6% of 310
3Areascrosses fields
Learn these firstLarge Language Model
Computed from the prerequisite graph, not assigned. How this works