Agent Fundamentals — Hostinger Agent

What Hostinger Agent is now

One product, merged from two: Kodee (dashboard support) + AI Agents (credits-based task agents).

Support Mode

What it's for: Account, hosting, website issues — troubleshooting, setup, billing
Cost: Always free
Availability: Keeps working even if Agent-mode credits run out

Agentic Mode

What it's for: Bigger, agentic asks: content, SEO plans, legal docs, strategy, scheduled jobs
Cost: Billed in credits, based on complexity/length
Availability: Paused once monthly credits are exhausted

Key alignment: customers don't choose the mode — the Agent decides automatically and will suggest switching if the other mode fits better.

Agent mode — quick contrast

How does Agent in agentic mode look for the customer?

Seamless experience — integrated UI
Request is business-related, not a Hostinger product question or issue
"Skills used" banner
Ambiguous requests are handled by the agent using context, and can switch modes as needed.

Skill generates an output; in the example, the outreach plan requested in markdown format.
It suggests next actions (that will require agentic skills) in line with the original request.
Credits are used upon task completion.

Agent mode — quick contrast

How does Agent in agentic mode look on HelpDesk?

Conversation view is similar for both modes
Main signals:
- is_premium=true on the premium_intent_agent
- There is a skill_invoked event

In support mode, is_premium = false

How a Conversation Actually Works

Message In

Customer or proactive trigger starts the chat

Routing

Product label assigned; re-checked on any turn

Base prompt

Initial behavioral guidelines and custom injections

Context

Past interactions, account attrs, up to 40 prior messages

Message out

Tool calling, function responses, detected language

Four sources it can pull an answer from

MCP Tool / Function

A wrong answer here is a debugging problem

Skills

Flexible, shaped by real conversation patterns

Knowledge

Public KB + internal raw text, knowledge-related queries that are not directly actionable.

Pure GPT

No hardcoded checks → more natural and engaging, but higher hallucination risk for concrete information.

The Handoff Ladder

When a customer asks for a human, the Agent escalates in steps.

Try to help

Answers directly, makes a light case for staying — once, at most

Offer a choice

Presents explicit buttons for the customer to decide

Loop in Guide Agent

A teammate feeds guidance behind the scenes

Hand off for real

Escalated to a human if needed/requested

Expected to skip the ladder when: a specialist already reviewed the case • a tool asked for a human directly • the message hit a legal/compliance keyword.

A real problem looks like the ladder skipping a step it shouldn't, or repeating one it already used.

Support mode

How does Agent in support mode look on HelpDesk?

is_premium=false on the premium_intent_agent
No skill_invoked event in support mode (if the agent switched it can appear in the same conversation)

Check a real example?

Which issues are worth digging deeper into?

Worth

Repetitive cases leading to:
- Handoff to a human
- Conversation crashes or errors
- Low CSAT
- Public complaints
Cases that are not repetitive but pose serious security, compliance, or reputational risks

Not worth

Isolated cases
Old cases (30 days or more)
Spikes that have already stopped
Issues known to be temporary, such as downtime
Low-impact cosmetic issues (formatting, wording, tone, occasional AI artifacts)
Cases with no identifiable root cause
Cases where no tools were used (pure GPT)
Apparent patterns where similar cases actually have different, unrelated root causes
Patterns based on weak or insufficient analysis, rather than consistent evidence

Where to look: HelpDesk event logs (tool_call, function_response, get_knowledge_agent, handoff) and the Observability tool (KodeeTrace) for base prompts and “Why?” analysis.

From one case to a pattern

One conversation shows what happened — never how widespread it is. Before pushing something up as an “Idea,” apply this bar.

50+

conversations / week

General baseline to justify deeper trend analysis. A practical threshold for effort, not a strict reporting rule.

…unless the impact is severe enough on its own:

Security
Incorrect policy guidance
Significant reputation or compliance risk
Severe customer impact (e.g. data loss)

A lower-volume problem can still deserve attention when the consequences are disproportionately important.