TELKOMSEL · NETWORK DIRECTORATE
AI Agents:
From Chatbots to
Getting Work Done
How autonomous agents actually work, what is already running in production, and how to start
Network Directorate · September 2026
5
active agents
27
use cases live
Rp0
local model cost
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 1
AGENDA
Five parts, one hour
Short on theory, then most of the time is spent watching the agents work
01
Concept
What an AI agent is, and how it differs from a chatbot
3 slides
02
Anatomy
Components, tool use, memory, model routing, platforms
5 slides
03
Evidence
The agent fleet and 27 use cases already in production
3 slides
04
Demo
Four live demonstrations in front of the room
4 slides
05
Closing
Security, governance, and how to get started
2 slides
Five agents working together from a single mini PC
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 2
CONCEPT
A chatbot answers. An agent delivers.
The difference is not how clever the model is — it is whether the model can act
CHATBOT (e.g. ChatGPT)
AI AGENT (autonomous)
System
Runs on the provider's cloud
Runs on our own hardware
Control
Limited to one chat window
Full — files, network, internal systems
Behaviour
Reactive: answers when asked
Autonomous and proactive: pursues a goal
Memory
Usually limited to one session
Long-term: recalls context and history
Action
Produces text only
Runs tools: terminal, files, APIs, schedules
The consequence: any task requiring several dependent steps can only be done by an agent, not a chatbot.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 3
ANATOMY
How an agent reaches the outside world
Users talk to it through ordinary messaging apps; behind that sits a gateway, a runtime, tools — and real connections to real services
CHANNELS
Telegram
Web browser
Command line
Control UI (dashboard)
Gateway
Routes every request, holds the session, and enforces who may talk to the agent.
Agent Runtime
Reasons, plans the steps, decides which tool to call, and checks the result.
TOOLS
Terminal, files, browser, schedules
MODEL
Local or cloud, chosen per task
What makes it persistent
Memory
Remembers preferences, conventions and past work between sessions.
Skills
85 ready-made capabilities, loaded on demand.
Schedules
12 jobs run on their own — daily, weekly, monthly.
Isolation
Each agent has its own memory, files and permissions.
CONNECTED SERVICES
live connections the agent uses today — not a roadmap
Google Workspace
Email, calendar meetings, Drive, Sheets
Instagram Graph
Reads insights, manages posts and reels
Telegram
Chat channel and message delivery
OpenAI
Image generation and heavy analysis
DeepSeek
Reasoning engine for four agents
Cloudflare
Free hosting and the web-chat model
Ollama
Local language model, no data leaves
ComfyUI
Four local image models, zero token cost
Second channel on the document agent
X / Twitter
Post search and publishing
Instagram token
Auto-refreshed weekly, 60-day expiry
GoBiz / GoFood
Credentials pending — API already researched
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 4
ANATOMY
The agent loop: perceive, think, act, check
An agent keeps cycling until the goal is met — that repetition is what separates it from a chatbot
Perceive
Reads the request and the current situation
Think
Decides what the next step should be
Act
Calls a tool or an application
Check
Verifies whether the goal is actually met
If the goal is not met, the loop repeats
Only when met → Result
A REAL RUN, STEP BY STEP
01
Read the file
The user sends a spreadsheet. The agent opens it and inspects the columns.
02
Analyse
It groups 159 transactions, computes totals, and checks every subtotal adds up.
03
Produce
It renders a one-page visual report at print resolution.
04
Verify
It opens its own output and re-reads it to confirm nothing is clipped or wrong.
05
Deliver
It sends the result back to the user in the same chat where the request arrived.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 5
ANATOMY
Memory: why an agent never starts from zero
Two kinds of memory work together — the conversation at hand, and permanent notes
Conversation memory
The session at hand
The message history of the current session, so the agent understands what has just been discussed without being re-briefed.
Long-term memory
Permanent notes
Deliberately stored facts: preferences, working conventions, environment details. They survive across sessions and across days.
REAL EXAMPLES FROM OUR SYSTEM
Document archive
The archiving agent files every incoming document into a catalogue holding date, category, summary and source.
Working conventions
The main agent remembers house style: colours, fonts, report layout and where files belong.
Measurable history
Every conversation is stored and can be counted — sessions, messages, even tokens per model.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 6
ANATOMY
Model routing: cloud, local, or specialist
This choice decides cost, speed and confidentiality — so we deliberately use all three
CLOUD MODEL
deepseek-v4-flash · 4 of 5 agents
Complex questions, long answers
Billed per token, only what is used
Best for: content, analysis, heavy reasoning
LOCAL MODEL
qwen3.8:27b · 1 of 5 agents
Runs fully on the mini PC we own
Zero token cost, no internet needed
Best for: archives, private documents
SPECIALIST TOOL
Claude Code · on demand
A dedicated coding and analysis agent
Used when the reasoning is genuinely hard
Best for: building apps, deep code analysis
pakisq is the safest agent in the fleet
It runs on a local model AND a locally-hosted agent platform. Nothing is sent to any external service and no data leaves the device — every document is opened, read and processed entirely on our own hardware. That is why confidential documents go to this agent.
01
Split by role
Agents handling private documents use the local model; agents needing speed use the cloud; specialist work goes to Claude Code.
02
Automatic fallback
Every agent has a backup engine. If the primary is unavailable, the agent switches without stopping.
03
Configuration, not code
Engines are settings, not source code. Adding or swapping one changes a single configuration line.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 7
ANATOMY
Choosing a platform: four families, many brands
Whatever the brand, the anatomy is the same — a model, tools, memory and a loop
SELF-HOSTED PLATFORMS
Runs on your own machine
Hermes
what we run today
OpenClaw
open-source assistant
n8n
visual, self-hostable
Dify
LLM app platform
AutoGPT
early autonomous agent
Full control, data never leaves. You own the updates and the uptime.
CODING & WORK AGENTS
Product you subscribe to
Claude Code
we use this too
ChatGPT Agent
OpenAI
Devin AI
Cognition
Perplexity Computer
research agent
OpenHands
open-source coding
Deepest at code and long tasks, but the vendor holds the runtime.
ENTERPRISE PLATFORMS
Vendor-hosted, governed
Microsoft Copilot Studio
Google Vertex AI Agent Builder
Azure AI Foundry
governed runtime
Salesforce Agentforce 360
AWS Bedrock Agents
Fastest path inside an existing cloud contract — data leaves the building.
FRAMEWORKS — BUILD IT
For developers, in code
LangGraph
stateful, auditable
CrewAI
role-based crews
OpenAI Agents SDK
GPT-centric
Google ADK
Gemini / Vertex native
Microsoft Agent Framework
AutoGen successor
Maximum flexibility, maximum engineering effort — you maintain it.
The practical question is not which brand wins — it is whether the data may leave your building, and who controls the tools the agent can reach.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 8
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 9
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 10
EVIDENCE
What the agents have actually produced
Every tile below is a real artefact from our own machine — screenshots, reports, posters and video
Live web application
pakis
9,066 stores, public address
Data report
pakis
159 transactions → one visual page
Market screening
pakis
Daily issuer scoring + trend chart
Social content
pakisem
Carousel built from own data
AI poster
pakissat
Text rendered locally, zero cost
Model benchmark
pakis
4 local image models compared
Presentation deck
pakis
This deck, built by the agent
Video reel
pakisem
Produced and published to Instagram
All of it produced by the agents themselves — including reading their own output to check it before delivery.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 11
EVIDENCE
What the agents actually write and publish
Real excerpts from files on our machine, plus video the agents produced end to end
ARCHIVE CATALOGUE
pakisq
| Date | Category | File |
| 2026-08-31 | Finance - Bank | indihome... |
| 2026-09-07 | Finance - Invest| lbs-market..|
| 2026-09-08 | Family | child-ac-log|
Every incoming document filed with a short
summary and its source, then searchable.
BUSINESS ACTION PLAN
pakissat
ONE-MINUTE SUMMARY
Two problems fixable this week, at no cost:
1. Not on Google → nobody searching
"siomay Margahayu" finds the shop.
2. GoFood rating 4.5 from 7 reviews, and 3
are complaints — below the 4.7 threshold
the algorithm favours.
WEEKLY PERFORMANCE REPORT
pakisem
7-DAY NUMBERS
Reach 227 · Views 860 · Likes 53 · Comments 4
Saves 0 · Shares 0 → saves/reach 0.00%
Interaction/reach 25.99% (healthy, but
nothing feeds distribution)
Reach −26% vs the 30-day average, likes +44%
→ "liked but not shared" keeps widening.
LIVE ANSWER, 0.4 SECONDS
pakis
Q: How many Indomaret stores are there,
and which province has the most?
A: There are 4,133 Indomaret stores. The
province with the highest total number
of stores is Jawa Barat (1,520 stores).
Answered live on the published dashboard by
a model running on Cloudflare.
VIDEO OUTPUT — INSTAGRAM REELS
The agent writes the script, lays out every frame, renders the video, and produces the caption and sound track.
Four reels are already published to the account.
FOUR REELS PRODUCED SO FAR
Cholesterol — 3.0 MB Insulin resistance — 2.8 MB
Dyslipidemia — 4.0 MB Cholesterol & exercise — 3.6 MB
Format 9:16. The frames shown here are taken straight from the rendered files.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 13
DEMO 1
Screening secondary-market issuers
The agent watches the market daily and tells us when something is worth a closer look
What you will see
The agent has already read a 45-page Islamic finance text and turned its investment standards into a checklist.
It then scores every issuer on the secondary market against those standards plus price, valuation and liquidity.
When an issuer passes the threshold, an alert arrives in chat without being asked.
How it runs
01
Read the standard
Islamic finance rules become explicit screening criteria.
02
Capture data
A daily snapshot of prices and demand.
03
Score
Every issuer ranked 0-100 on the same criteria each day.
04
Alert
A short summary plus any issuer worth attention.
Worth noting
The investment framework comes from Ust. Ammi Nur Baits' book on Islamic transactions — real assets, no interest, no uncertainty, no guaranteed returns.
The platform itself was founded by and is supervised by Ust. Erwandi Tarmizi, from the same school of Islamic commercial law.
The score is a screening aid, not a buy recommendation.
FALLBACK PLAN
Historical snapshots and trend charts are already stored, so the demonstration still works if the live feed is unavailable.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 13
DEMO 2
Building an app, then publishing it
The application is built in Claude Code; our agent is what puts it on the internet
LIVE NOW
minimarket-indonesia.pages.dev
9,066 minimarket stores mapped across Indonesia.
Answered 200 OK from anywhere in the world.
Hosting cost: zero. Bandwidth: uncapped.
Stays online even when our mini PC is off.
HOW IT RUNS
01 Build in Claude Code, tested locally
02 Prepare files so no server is needed
03 The agent deploys and verifies data loads
04 A public address is handed back to us
FALLBACK PLAN
The application is already published and reachable, so this demonstration does not depend on the office network or on our mini PC being switched on.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 14
DEMO 3
Social content that closes the loop
The agent writes content from the account's own history, then measures what happened
What you will see
The agent drafts carousels and reels based on what has actually been shared most from the account.
Once published, it pulls real performance data: reach, saves and shares per post.
The audience sees next week's recommendation being built from numbers, not opinion.
How it runs
01
Read history
Pull performance of previous posts.
02
Write
Decide the topic, slide order and text for each slide.
03
Design
Lay out the slides in the owner's visual identity.
04
Measure
Pull insights and produce the next recommendation.
Worth noting
Health claims are worded carefully so they stay medically safe.
Publishing to Instagram is explained verbally rather than demonstrated.
Performance data comes from the official API, not from screenshots.
FALLBACK PLAN
All content files and insight reports are stored locally, so the demonstration still runs even if the API quota is exhausted.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 15
DEMO 4
From a hand sketch to a corporate deck
Hand over a sketch on paper; the agent turns it into a finished presentation
What you will see
An audience member is asked to sketch an idea on paper and photograph it.
The agent reads the sketch, understands the structure, and builds a deck in the established corporate identity.
Charts and diagrams are redrawn as vectors, so everything stays sharp.
How it runs
01
Read the sketch
Recognise the intent, order and relationships.
02
Structure
Lay out the slides and the weight of each section.
03
Draw
Produce design, charts and illustrations to the identity.
04
Check
Verify: no clipped text, no overlapping elements.
Worth noting
The deck you are looking at now was made with exactly this flow.
One font family and one colour scheme, with no leftovers from foreign templates.
Visual quality matters more than the speed of generation.
FALLBACK PLAN
The deck currently open in front of the audience is its own proof — nothing needs to be re-generated during the demonstration.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 16
CLOSING
Security and governance
Four rules we have held since day one — every one of them learned from real experience
01
Tokens, never passwords
The agent never holds a password. Every integration uses a scoped token that can be revoked without changing anyone's credentials.
Permissions limited to what the task needs
Stored separately with strict file permissions
Short-lived and refreshed automatically
02
Least privilege
Each integration is given the narrowest access that still works. If it only needs to read, it does not get write access.
Write access only when genuinely required
Each token scoped to a single service
Revoking one token does not disturb the rest
03
Agent isolation
Every agent lives in its own space: its own memory, files and permissions. One agent cannot read another agent's data.
No mixing of working context
Conversations do not leak between agents
Explicit rules on what must never be disclosed
04
An auditable trail
Every conversation, model call and produced artefact is recorded and can be recounted at any time.
Sessions, messages and tokens are counted
Files archived in a structured catalogue
Can trace who asked for what, and when
THE ONE RULE THAT MATTERS MOST
Give an agent access that can be withdrawn, not trust that cannot be recalled.
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 17
CLOSING
When to use an agent, and how to start
Three small steps to try it without disturbing the work already in flight
01
Start with one task
Pick a single repetitive task with a clear beginning and a clear result. Do not start from your most complicated process.
Example: producing a recurring report from data you already have.
02
Give it a home
You need one machine that can stay switched on — a small computer or a server — plus one language-model account. No special hardware.
Example: the mini PC we already use, with nothing added.
03
Add capabilities
Once one task runs reliably, add the next one. Each new capability is stored separately, so nothing already working is disturbed.
Example: from producing a report, to watching the market and sending alerts.
Thank you
Questions, an idea for a task you would like to try, or a repeat of any demonstration — let us discuss it now.
Q & A
Telkomsel · AI Agent Training
Network Directorate · Sept 2026 · 18