AI Agents and Agentic AI
An AI agent is a system in which a language model decides, step by step, which actions to take (call a tool, read a file, browse, delegate) to reach a goal, observing the results and looping until done. Agentic AI is the broader label for software built around such loops. The core pattern is the agent loop: model proposes an action, the surrounding code executes it, the result is fed back, repeat. The pattern was formalised in the ReAct paper (Yao et al., submitted 2022-10-06), which interleaves reasoning and tool actions.
Anthropic’s “Building effective agents” (2024-12-19) draws a useful line: workflows are LLM calls orchestrated along predefined code paths, while agents let the model direct its own process and tool use. Start with the simplest thing that works; add autonomy only where the task needs it.
Levels of autonomy (rule of thumb)
- Single LLM call with retrieval or a tool (assistant).
- Fixed workflow / chain of calls (prompt chaining, routing, parallelisation).
- Agent loop with human approval on risky steps.
- Long-running autonomous agent, often with sandbox, memory and sub-agents.
How it is used in 2026
Coding agents, browser/computer-use agents and personal assistants all share the loop; what differs is the harness (tools, permissions, context management) around the model. Interfaces between parts are being standardised: MCP for tools, A2A for agent-to-agent calls, AGENTS.md for project instructions. Reliability, security (prompt injection) and evaluation (agent-evals-and-observability) are the main open problems.
Related notes
autonomous-agents, agentic-development, multi-agent-systems, agent-orchestration-overview, task-decomposition-patterns, agent-memory-systems, computer-use-agents, agent-sandboxes, claude-agent-sdk, openai-agents-sdk, langgraph, how-to-build-reliable-ai-agent-systems-for-production, reasoning-models.
Sources
- https://arxiv.org/abs/2210.03629 (ReAct: Synergizing Reasoning and Acting in Language Models, v1 2022-10-06, re-checked 2026-10-07)
- https://www.anthropic.com/engineering/building-effective-agents (Anthropic, 2024-12-19; workflow/agent definitions re-checked 2026-10-07)
Open items
- Autonomy-level taxonomy above is an editorial rule of thumb, not a cited standard.