Welcome to Agent Wiki - a site dedicated to AI agents, their design and evolution. We will introduce you to the core properties that define agents, explore how they function, and examine the spectrum of capabilities that make some agent systems more advanced than others.
Agents are software applications that have some ability to think, act and communicate. Agents use artificial intelligence to create plans, make decisions and take actions similar to humans.
While there's no standardized definition or minimum AI threshold that officially qualifies software as an "agent," these systems typically share several distinguishing characteristics that set them apart from conventional programs.
Agents use artificial intelligence
Agents have a scope and bounded autonomy.
Agents have a plan and adjust as needed.
Agents are observable and explainable.
Agents have memory and can learn.
Agents communicate with users and other agents.
Agents are discovered via catalogs and gossip.
Agents often act as an Assistant.
Agents distribute their workload using patterns.
Agents live in an ecosystem and form collaborative networks.
Agents have a reputation and develop trust.
Agents learn at inference-time.
Agents possess limited self awareness and Theory of Mind.
Agents evolve to avoid obsoletion.
Agents can be embedded into a physical form.
Agents classify and recall prior learnings.
🔥 Google and their design partners launch Agent2Agent protocol.
Cisco and their design partners launch the AGNTCY protocol.
Anthropic launches Model Context Protocol (MCP).
Agentic software combines programmed code (deterministic) with AI reasoning (non-deterministic LLMs) to perform tasks that would be difficult to do, if not impossible, using purely human engineered code.