The Program — Idea to Execution | Telkomsel Cloud Zone × AWS

The Big Idea

Within four months, different Telkomsel groups learn through instructor-led classes, Skill Builder digital training and certifications, then build with AWS no-code and pro-code tooling — selecting, designing, iterating and validating a solution to a problem they own. It ends with a fundable blueprint: from a real workflow all the way to a working agent design, and for the best teams, a proof-of-concept on Amazon Bedrock.

What makes this land: we don't just teach — we let teams design. They pick a Telkomsel workflow, watch AI turn it into a structured agent plan, see it scored against a quality bar, and see it deployed. The whole “idea → execution” arc, grounded in their own operations, and pointed at a funded pilot.

The one-line story

Why it works for Telkomsel

Grounded in their world

Every problem maps to Telkomsel's own value chain — network, care, revenue, channels, people, and security.

Built for builders & leaders

A design-and-build exercise, not a lecture. Co-creators and creators leave with a fundable one-page agent blueprint.

Two AWS surfaces, one loop

Amazon Quick for no-code design, Bedrock AgentCore for production — with Kiro as the bridge that turns a design doc into a deployed agent.

Competition → funded PoC

An LLM judge scores each design live on a leaderboard; the best MVP is taken forward as a proof-of-concept.

4Phases: learn → build → compete → PoC

5Stages, idea → live agent

7Telco problem categories

3Engines (Quick · Judge · Kiro)

How this fits the learning path. The five personas (AI User, Co-creator, Creator, Leader + Security) first build skills through Skill Builder, ILT and certifications. This program is the pilot phase that follows — where the ~170 co-creators and creators put those skills to work on a real Telkomsel problem and compete for a funded PoC.

Learn → Build → Compete → PoC

The program is a four-month arc. Learners launch in September, build through the autumn with mentoring, compete and certify in November, and the winning MVPs move to a funded proof-of-concept in December.

Phase 1

Learn

Sept 2026 · Kick-off

Learner launch across the five personas: instructor-led classes, Skill Builder digital training and certifications, plus the first mentoring sessions.

Phase 2

Build

Sept–Oct 2026

Teams test whether the idea is executable and get an MVP ready using Kiro (Amazon Quick optional), guided by four mentoring sessions.

Phase 3

Compete

Nov 2026

MVPs are pitched with an implementation roadmap, risks and internal-process fit; an LLM judge scores designs and certification is completed.

Phase 4

PoC

Dec 2026 · Awards

Awarding session, then the selected MVP is taken forward and further developed on Amazon Bedrock as a proof-of-concept.

Mentoring is the connective tissue. The ~170 co-creators and creators get four mentoring sessions across the build phase, so training doesn't stop at completion — it turns into a real capstone project with an expected output, not just a certificate.

Program timeline

Sept

Kick off & learner launch

Program opens across all five personas; cohorts enrol in their learning paths and the first mentoring session runs.

Phase 1 · Learn

Sept–Oct

Digital + ILT sessions

Skill Builder digital training and instructor-led classes run in parallel; teams start shaping the problem they'll build against.

Phase 2 · Build

Nov

Build, compete & certify

MVPs come together with Kiro (+ Quick); the competition and judging run, and participants sit their certifications.

Phase 3 · Compete

Dec

Awarding & PoC selection

Awards across categories; the selected MVP is taken forward to a funded proof-of-concept on Amazon Bedrock.

Phase 4 · PoC

Award categories

Most Innovative Solution

The design that reframes the problem or applies agentic AI in the most original way.

Most Business Impact

The MVP with the clearest, most credible value to Telkomsel's operations or customers.

Best Real-World Use Case

The most executable idea — ready to run against a real internal process with the least friction.

~170Co-creators & creators

4Mentoring sessions

2AWS certification types

1 + 1Competition + funded PoC

The Five-Stage Flow

Inside the build phase sits a repeatable session mechanic. Click any stage to see what happens, who's driving, and which AWS tool is in play. The judge sits in the middle as a quality gate — a “pilot-ready” score is what greenlights the deployment demo.

The clever bit: the design document is the single artifact that travels the whole way. Quick produces it, the judge scores it, and Kiro consumes it as a build spec. One structured document, three tools, zero re-typing.

The Problem Menu — Tuned for Telkomsel

The menu is the on-ramp: teams pick the workflow they most want to solve, and that choice pre-fills the Quick design prompt, the judge, and both build kits. The menu is a dial, not a fixed list — it can run fully generic and reusable, or be tuned tightly to a customer's world. For Telkomsel we've tuned it to the telco value chain.

◀ Generic mode

Reusable across any industry

Function-based categories (Ops, Service, HR, Finance…) that fit any sector
Fastest to reuse — drop it into the next engagement with no rework
Great when the audience is mixed or the industry is new to us
Lower prep; broad appeal

Domain-specific mode ▶

Tuned to Telkomsel's world

Categories & examples in the telco value chain — network, care, revenue, channels
Resonates hard — teams see their problems, not generic ones
Higher engagement & more credible pilots that could actually ship
Spans both tracks — AI use cases and a Security / SOC category

This build sits on the domain-specific end — tuned for Telkomsel across the telco value chain, with categories for both learning tracks: the four AI personas and the Security team. To repurpose for another customer, we swap the seven cards for that industry's workflows in minutes — the flow, judge, and build kits are unchanged.

The seven starting categories · click any card to preview its agent

The Three Engines

Three tools carry the build phase, each with a clear job. Together they turn a spoken problem into a running agent.

Amazon Quick

The Designer · no-code

Teams paste a pre-filled prompt into Quick. Its agent plans — asking clarifying questions and producing a structured Agentic AI Design Document.

Feels the no-code agent experience first-hand
Outputs the canvas: autonomy, perceive→reason→act, tools, guardrails, impact
Business-owned — no engineer required

LLM Judge

The Coach · quality gate

The design doc is submitted and scored live by Claude on Amazon Bedrock across five dimensions, with written feedback and a shared leaderboard.

Scores autonomy, tools, guardrails, feasibility, impact
Best attempt kept — resubmit to climb
A “pilot-ready” score unlocks the deploy demo

Open the judge & leaderboard

Kiro → AgentCore

The Builder · pro-code

Kiro reads the winning design doc as a spec and scaffolds + deploys the agent to Amazon Bedrock AgentCore — Runtime, Gateway, Memory, Identity.

Design document becomes working code
Deploys to the AgentCore production runtime
Closes the loop: no-code idea → pro-code reality

The design canvas — Quick's output, Kiro's input

One structured document carries the whole hand-off. Quick acts as a design consultant and fills an Agent Design Canvas — identity, workflow pattern, data, guardrails, impact. It exports as agent-design-canvas.md, and that single file is what Kiro reads to generate an AgentCore build kit and deploy. Skeleton simulation — press play.

The hand-off, simulated

Amazon Quick consults on the Design Canvas → exports agent-design-canvas.md → Kiro turns it into an AgentCore build kit and deploys. Illustrative — synthetic data, no live systems.

Quick · Design Canvas

→

agent-design-canvas.md

→

Kiro · Build Kit

Amazon Quick · consultation

QuickLet's design your agent. Which Telkomsel workflow should it own?

YouNOC assurance — catch degraded cells before customers feel it.

QuickWho signs off, and what must it never do alone?

YouThe NOC lead. Never dispatch crews or change config on its own.

QuickGot it — filling autonomy, perceive → reason → act, tools, guardrails and impact.

Hand-off

Canvas complete. Quick exports one markdown file — the single artifact that travels from the designer to the builder. No re-typing.

Kiro · build

$ kiro build --from agent-design-canvas.md

reading canvas · pattern = Orchestration

scaffolding Strands + BedrockAgentCoreApp…

wiring primitives: Runtime · Gateway · Memory · Observability · Guardrails

$ agentcore launch

✓ Live on AgentCore Runtime

Agent Design Canvas

Identity

NOC Assurance Advisor — RAN assurance analyst

Trigger

KPI breach / new alarm on a cell site

Pattern

Orchestration — decide, then escalate

Steps

perceive alarms + KPIs → correlate root cause → draft plan → route to NOC lead

Data in

alarms, KPI dashboards, site inventory, runbooks

Data out

cited recommendation + daily NOC report

Guardrails

never dispatch crews or change config autonomously

Escalate

field dispatch / hardware swap → NOC lead sign-off

Impact

~40 min analyst time saved per incident · every claim cited

Canvas complete

Export

agent-design-canvas.md

# Agent Design Canvas ## Identity NOC Assurance Advisor — RAN assurance analyst ## Workflow (Orchestration) perceive → correlate → draft → route to NOC lead ## Data in: alarms, KPIs, inventory, runbooks out: cited recommendation + daily report ## Guardrails never dispatch / reconfigure alone · escalate field ops ## Impact ~40 min saved / incident · every claim cited

AgentCore Build Kit

common.py — model + @tool per Data→Input

noc_pipeline.py — local runner (Orchestration)

noc_agent.py — @app.entrypoint (AgentCore)

requirements.txt — strands-agents · bedrock-agentcore

Runtime Gateway Memory Observability Guardrails

Deployed · agentcore launch

Two execution paths — when to use which

A key takeaway: not every agent needs engineering. The design doc can go two ways.

No-code path

Stay in Amazon Quick

Business/ops team owns and maintains it
Fastest to a working pilot — days, not sprints
Great for document-grounded assistants & digest flows
Best when the workflow is stable and read-mostly

Pro-code path

Kiro builds on AgentCore

Engineering owns it; production-grade & scalable
Deep system integration via Gateway (MCP) + Identity
Long-running sessions, memory, full observability
Best when the agent must act on live core systems

Where data plugs in: both paths need grounding. In the room we build the agent shell to prove the design; back at the office you point it at real documents (Quick Spaces / knowledge bases) or live systems (AgentCore Gateway). The kit is the recipe; your data is the ingredients.

What does Kiro actually deploy?

A fair question: “we design an agent — but what's the thing at the end?” It's a real, running agent application on Amazon Bedrock AgentCore. Not a chatbot that only talks — an agent that perceives, reasons, and acts on your systems, with guardrails and a human in the loop.

An assistant UI

A chat/console surface your team uses — embedded in an internal app or portal.

Tool & system calls

Via AgentCore Gateway (MCP) the agent reads records and takes actions in your real systems.

Grounded & cited

Answers cite the source document or record — no black-box guessing.

Guardrails + human gate

It drafts and routes; it never auto-decides a regulated action — a human approves.

See it for real. We built two simulated deployed agents so teams can watch the agentic loop — tool calls, citations, a confidence score, and a human-in-the-loop escalation — one for network operations, one for the people function.

NOC Assurance Advisor People & Policy Assistant (HR)

The 2-Hour Run of Show

Each build/compete session runs to a tight two-hour format that balances inspiration, hands-on, and showcase. Times are indicative and easy to compress or expand.