Ask a chatbot about an order and it may explain where to find its status. An agent can open the relevant system, retrieve the status and prepare the next step.
This is a general example, not a claim about a specific Ignition feature. The key difference is that an agent does not only work with text. It receives access to clearly defined tools and performs steps within a controlled workflow.
What is a chatbot and what is an agent
Chatbot
holds a conversation;
answers questions;
explains or generates content.
Agent
understands a task and its context;
selects an allowed tool;
retrieves necessary data;
performs a specific action;
checks the result;
continues the workflow according to defined rules.
An agent does not need unlimited autonomy. Its tools, permissions, approval steps and system rules define its scope.
Basic architecture
The user provides a goal through an interface such as voice or Telegram. The orchestrator or agent recognises the intent and chooses an allowed tool. The tool performs a narrow operation against an API or internal system. The result returns to the agent for checking and either a response or the next allowed step.
The interface is not the agent, and the model is not the entire system. It is one layer surrounded by integrations, state, rules and controls.
Voice is just another interface
A voice flow can look like this:
Voice → speech-to-text → agent → tool or workflow → response → text-to-speech.
An agentic system does not need its own app. The same backend can be available through voice, Telegram or another messaging interface. Voice and Telegram are confirmed in Ignition's internal stack; Messenger is therefore not listed as an active channel.
One agent versus specialised agents
A simple process may need one agent with a few tools. In a broader system, an orchestrator can choose a specialised agent, which uses its tools and returns the result to the user or the next step.
Specialisation can simplify permissions and instructions. A multi-agent design is not automatically better: it adds coordination, state and more failure points. It is useful when clear roles produce more predictable behaviour than one general agent.
What we actually use at Ignition
At Ignition, we use our own agents and workflows, including Hermes. We build the orchestration, task routing, integrations and automation logic. Voice and Telegram provide interfaces to connected workflows.
We use AI Agent Infrastructure internally for:
voice input and output;
Telegram interaction;
tools and external APIs;
task routing between agents;
automated briefings;
scheduled tasks;
connections to internal systems.
Hermes is our internal AI agent. We use and integrate OpenClaw as a third-party/open-source component. We build integrations on top of Paperclip, also a third-party/open-source component. Foundation models and external AI APIs are third-party technologies, not Ignition products.
Where an agent can fail
An agent can choose the wrong tool, send invalid arguments or rely on stale information. A model can hallucinate a plausible but incorrect response. During a longer process it can lose context or misread the current state.
The greatest risk arises when a wrong answer triggers an unwanted action. Drafting an email is different from sending it; finding an order differs from changing its status; preparing a payment differs from confirming it.
Why an agent needs controls
Permissions define which data and actions are available.
Approval steps stop sensitive actions and ask a person to confirm.
Logs record the tool, input and result.
Tool restrictions expose only the operations needed for a role.
Explicit state records where the workflow is instead of relying only on conversation history.
Human-in-the-loop keeps financial, legal or operational decisions with people.
These controls do not create absolute safety. They reduce risk and make failures easier to catch, explain and correct. Design should consider the impact of a wrong action, not only whether an agent can answer successfully in a demo.
For the distinction between AI tools and vendor responsibility, read App development with AI: when it makes sense and what to watch out for.
Conclusion
The most interesting part of AI agents is not the conversation. It is the safe connection between a model and the systems in which it can perform real work. Permissions, tools, state and control points determine whether the result is a useful workflow or only an impressive demo.
