Agentic AI

General definition

Agentic AI is a design paradigm, not a specific model or product. A system is agentic when it can decompose a high-level objective into sub-tasks, choose which tools or sub-agents to invoke at each step, observe the outcomes, and revise its plan accordingly, all without a human approving each action. The degree of autonomy varies: some systems ask for confirmation before consequential steps; others run end-to-end loops until a stopping condition is met.

• Planning - breaking a goal into an ordered sequence of steps
• Tool use - calling functions such as web search, code execution or database writes
• Observation - reading the output of each action and deciding what to do next
• Self-correction - detecting failures or unexpected results and adjusting the plan
• Multi-agent coordination - delegating sub-tasks to specialised agents and aggregating results

Agentic patterns introduce new risks compared with single-turn LLM calls. Mistakes compound across steps, prompt injection can hijack a running loop, and an agent with write access to external systems can cause hard-to-reverse side effects. Responsible deployments include human-in-the-loop checkpoints, rate limiting on tool calls, and careful scoping of what an agent is permitted to do.

In the Ethora ecosystem

Ethora’s AI SDK supports agentic patterns inside a real-time messaging environment. You can wire an agent to a chat room so it monitors incoming messages, decides when to act, calls your business APIs via tool calling, and posts results back as a participant in the thread. Multiple specialised agents can be deployed in the same app, each scoped to a different room or topic.

For teams that need full control over agentic behaviour, the self-hosted LLM agent configuration runs the entire loop on your infrastructure. No user messages, no tool outputs, and no intermediate reasoning traces leave your environment, which is a hard requirement in many compliance frameworks.