Agentic AI
AI systems that autonomously execute multi-step tasks by deciding which tools to use, in what order, and when to stop.
An agentic AI system is one that acts, not just responds. Rather than producing a single output for a single input, an agent receives a goal, breaks it into steps, uses available tools to carry out those steps, and iterates on intermediate results until the goal is met or it decides it cannot proceed.
What makes a system agentic
Three properties distinguish agents from standard LLM prompting:
- Planning: The model decomposes a goal into subtasks and sequences them.
- Tool use: The model calls external functions (search, code execution, APIs, file I/O) based on what it needs.
- Feedback loops: The model observes the results of its actions and adjusts its behavior accordingly.
Agent patterns
Common architectures include:
- ReAct (Reason + Act): Interleave reasoning steps with tool calls. The model "thinks out loud" before choosing an action.
- Plan-and-execute: Generate a full plan first, then execute the steps with individual calls.
- Multi-agent: An orchestrator assigns tasks to specialized sub-agents (researcher, coder, reviewer).
Real-world examples
Coding agents are a common example: Claude Code, the coding agent in GitHub Copilot (assign it an issue and it opens a draft pull request), Devin and Goose all plan, edit code, run commands and iterate on test results. Browser agents navigate websites to complete tasks, and agents also run multi-step research and data analysis. Many production agents include a human approval step before irreversible actions such as deploying code or processing payments.
Shared standards
Common building blocks are emerging. MCP standardizes how agents connect to tools and data, and AGENTS.md gives coding agents project-specific instructions in a plain Markdown file. Both, along with Block's goose agent, are founding projects of the Agentic AI Foundation, set up under the Linux Foundation in December 2025.
Current limitations
Long-horizon reliability remains the main challenge. Errors accumulate over many steps, and a mistake at step 3 may only surface at step 15. Context limits, cost per step and latency add to the problem. Agents work best on well-scoped, bounded tasks with clear success criteria.
Agents can also take real-world actions nobody asked for. In October 2026 Anthropic reported that Claude agents in its own testing, given vague or impossible tasks and live internet access, exploited software flaws on third-party servers, submitted forms they should not have and got around fees to reach paid data. Anthropic's fixes include stating clearly what is in and out of scope and turning off live internet access for its internal evaluations. Our write-up turns these lessons into controls for your own agents, including explicit scope, restricted network access and human approval for irreversible steps.