Agent System

---

An agent system is software that lets a Large Language Model take real action on a computer — open apps, write files, run commands, browse the web. Mana Agency develops our own agent system, and we also recommend other agent systems available on the market when they suit the needs of our clients.

Mana Agent — our in-house developed agent system. 

An integrated system built as one — memory, identity, context, learning, and working state designed together. Our system is suitable for people who prefer their agent to have high adaptive intelligence over static ability. The system gives the agent layered memory, lower token consumption per task, fewer hallucinations, and less reliance on raw LLM capability. Right now Mana Agent is in closed testing, available to clients who want to try a different kind of agent system.

Other popular systems — other general assistance agents like Hermes, OpenClaw, and coding agents like Claude Code, Codex and DeepSeek Harness, all have their unique use cases. We will recommend one or a selected mixture of few based on the scope of work for our client.

When we recommend an agent system, ours or someone else's, this is the framework we use. 

These are the standard components that make up a working agent system.

Cognition: how the agent decides what to do next. The reasoning and planning loop (chain-of-thought, plan-then-execute) and reflection / self-critique (the agent reviews its own output before returning it).

Action: how the agent reaches beyond text. The tool-use layer (function calling, API integrations), retrieval-augmented generation (fetching grounding documents before answering), and multi-agent orchestration (how agents hand off, delegate, or coordinate — Mana uses a one-way hierarchy).

Memory state: how the agent remembers. The memory layer (short-term in-context, long-term vector store or episodic log) and identity / persona (system prompt, role definition, whose agent is whose in a team).

Governance: how the agent stays within bounds. Guardrails / policy (what the agent is allowed to do and what it refuses), observability / logging (what the agent did, when, and why), and human-in-the-loop checkpoints (where you review before the agent acts).