AAP / USE / 01
AI Tools Mastery
ChatGPT, Claude and automation, used with judgement.
Open the syllabusPeru · AI agents at work · Adults 18 and over
A demonstration agent works on the first try. A working agent has to survive Monday morning: a supplier's system that answers slowly, a payment service that times out, a request that may or may not have gone through. This page is for professionals in Peru who want to build or manage AI agents that hold up in real business processes. It covers what agents actually do in companies, the reliability habits engineers have learned the hard way, what Peru's 2025 AI regulation sets out for private companies, and one lesson in which an agent bills a client twice.
Live classes since 2020 · 10,000+ students worldwide · live, online and taught in English
In brief
Modern Age Coders teaches AI agents to working professionals in Peru: live online classes, taught in English, for adults who want to build agents that handle real business tasks. You learn low-code agents in Microsoft Copilot Studio or Google Gemini Enterprise, or coded agents in Python, and the practices that make them trustworthy: timeouts, safe retries, protection against duplicate actions, and handing a case to a person when the agent should stop. Mentors in India teach one to one or in a group of fifteen to twenty, at hours that fit your working week in Peru.
Pick your platform
Choose by the tools your company already uses. All three are for adults, and the first two are taught one to one.

AAP / 01
For teams on Microsoft 365. Microsoft describes Copilot Studio as a "graphical, low-code studio for building and managing AI-powered agents and workflows". One to one, built around your own processes.
Open the syllabus →
AAP / 02
For companies on Google Workspace and Google Cloud. A one-to-one course on designing, connecting and supervising agents on Google's enterprise platform.
Open the syllabus →
AAP / 03
For professionals who want to code their own agents and automations: Python scripts, APIs and AI tools, with the retry and timeout patterns from the lesson below built in.
Open the syllabus →Agents at work
Take a defined task from start to finish: read a request, look things up, act in another system and report back. OpenAI's guide puts it in one line: "Agents are systems that independently accomplish tasks on your behalf."
A customer asks about an order; the agent checks the order system, sees the delay, and drafts an update with the new date. A person approves anything that commits the company.
An agent gathers a supplier's invoice, matches it with the purchase order and the delivery note, and files it, or flags the mismatch for someone in finance.
An agent watches a shared mailbox or a dashboard, sorts what arrives, and alerts the right team, leaving a clear record of what it saw and why it acted.
In each case the agent calls other systems: an order database, an accounting platform, an email server. Every one of those calls can be slow or fail, and some of them change things in the world. The quality of an agent at work depends less on how clever its language model is than on how it behaves when a call goes wrong. That is the part most short courses skip, and the part our course teaches first.
Developers who already use agents report real but uneven benefits. In the Stack Overflow Developer Survey 2025, about 70% of agent users agreed agents had reduced the time spent on specific development tasks and 69% that they had increased productivity, but only 17% agreed agents had improved collaboration within their team. Individual speed is the easy win; fitting an agent into a team's processes is the harder work.
When tools fail
The agent does not know whether the action happened. A timeout means no answer arrived, not that nothing was done. Engineers at Amazon have written about exactly this problem for years.
| Idea | In AWS's words | For an agent at work |
|---|---|---|
| Timeouts | "Timeouts keep systems from hanging unreasonably long" | Every tool call needs a limit, or the agent waits forever |
| Unknown outcome | "A timeout or failure doesn't necessarily mean that side effects haven't happened." | An invoice may exist even though the call "failed" |
| Idempotency | Designing APIs "to be idempotent, meaning they can be safely retried" | Send the same request key so a repeat does not repeat the action |
| Backoff and jitter | "Jitter adds some amount of randomness to the backoff to spread the retries around in time." | Wait longer each time, by a random amount, and stop after a few tries |
Source: Amazon Web Services, Builders' Library, "Timeouts, retries, and backoff with jitter", published 5 December 2019, read on 9 November 2026. The right-hand column is our own reading for agents.
AWS describes the most common pattern as exponential backoff, where the wait grows after every attempt, usually with a ceiling: "This is called, predictably, capped exponential backoff." Its next step is to limit how many times a client retries at all.
The article adds a line worth remembering: "retries are selfish". Each retry demands more work from a system that may already be struggling. An agent that retries instantly and endlessly can turn a small outage into a large one.
OpenAI's agent guide lists "Exceeding failure thresholds: Set limits on agent retries or actions" among the moments when control should pass to a human. A well-designed agent stops, says clearly what it tried and what it does not know, and hands the case over.
Peru's 2025 AI regulation also names human supervision as a principle (article 7 h of DS 115-2025-PCM). We mention it as context; the engineering reason alone is enough to build the hand-off.
A real lesson
This lesson comes from our professionals' course. An agent creates an invoice for a small bakery. The billing service does the work, but its reply is lost. What the agent does next decides whether the client gets one invoice or two.
import random
class InvoiceServer:
"""Invented billing service: the first call is processed but its reply is lost."""
def __init__(self):
self.invoices, self.seen_keys, self.calls = [], {}, 0
def create_invoice(self, client, amount, key=None):
self.calls += 1
if key in self.seen_keys:
return self.seen_keys[key] # same key: same invoice, no new one
number = "F-%03d" % (len(self.invoices) + 1)
self.invoices.append((number, client, amount))
if key:
self.seen_keys[key] = number
if self.calls == 1:
raise TimeoutError("no reply within 2 s") # done, but the agent never hears
return number
def naive(server):
for attempt in range(1, 4):
try:
return server.create_invoice("Panaderia Sol", 120)
except TimeoutError:
pass # retry at once, no key
def careful(server, max_attempts=4, base=1.0, cap=8.0, rng=random.Random(7)):
key = "order-2026-118" # one key per real-world order
for attempt in range(1, max_attempts + 1):
try:
return server.create_invoice("Panaderia Sol", 120, key=key)
except TimeoutError:
if attempt == max_attempts:
break
wait = rng.uniform(0, min(cap, base * 2 ** attempt)) # capped backoff, full jitter
print(" attempt", attempt, "timed out; would wait", round(wait, 2), "s") # printed, not slept
return "HANDED TO A PERSON"
a = InvoiceServer(); print("naive ->", naive(a), "| invoices:", [i[0] for i in a.invoices])
b = InvoiceServer(); print("careful->", careful(b), "| invoices:", [i[0] for i in b.invoices])
class DownServer(InvoiceServer):
def create_invoice(self, client, amount, key=None):
raise TimeoutError("service down")
print("down ->", careful(DownServer()))The service, client name, order and amount are invented by us. Waits are printed rather than slept so the lesson runs instantly; the random generator is seeded so the output is the same every time.
naive -> F-002 | invoices: ['F-001', 'F-002'] attempt 1 timed out; would wait 0.65 s careful-> F-001 | invoices: ['F-001'] attempt 1 timed out; would wait 0.3 s attempt 2 timed out; would wait 2.6 s attempt 3 timed out; would wait 0.58 s down -> HANDED TO A PERSON
Naive agent: two invoices. Careful agent: one invoice, after one wait. Against a service that is down, the careful agent tries four times, waits three times, and hands over.
The naive agent did not malfunction. It retried, which is usually sensible. The damage came from retrying an action with a side effect without any way for the other system to recognise a repeat. AWS states the rule plainly: APIs with side effects "aren't safe to retry unless they provide idempotency".
In a real company the "invoice" might be a payment, a purchase order, a message to a customer or a change to stock levels. The pattern is the same, and so is the cost of getting it wrong.
Many business APIs accept an idempotency key or a client reference for this reason; some do not, and then the agent must check whether the action happened before trying again. Low-code platforms such as Copilot Studio hide some of this, which makes it more important that the person designing the agent knows what to ask.
The waits in the output (0.3, 2.6 and 0.58 seconds) look irregular on purpose. That is the jitter: if a hundred agents retry against the same struggling service, random waits stop them all arriving at the same instant.
max_attempts and cap. What is the longest the agent could wait before handing over, and is that acceptable for this task?Peru timetable
A phased timetable by sector. DS 115-2025-PCM, which regulates Law 31814 and was published in El Peruano on Tuesday 9 September 2025, gives private developers and implementers between one and four years, by sector, to implement article 25 and chapter II of title VI.
| Uses of AI in | Period in the regulation | Which ends around (our arithmetic) |
|---|---|---|
| Health, education, justice, security, economy and finance | One year | September 2026 |
| Transport, commerce and labour | Two years | September 2027 |
| Production, agriculture, energy and mining | Three years | September 2028 |
| All other uses | Four years | September 2029 |
| Small companies, annual sales above 150 and up to 1,700 UIT | Two years | September 2027 |
| Micro companies, annual sales up to 150 UIT | Three years | September 2028 |
Source: DS 115-2025-PCM, Reglamento de la Ley 31814, first complementary final provision, El Peruano, 9 September 2025; each period runs from the day after publication. The right-hand column is our own calendar arithmetic, not a legal interpretation. Read on 9 November 2026.
This is published context, not legal advice: which period applies to a given company, and exactly what it must implement, is a question for its lawyers. What a course can do is make sure the people building and running agents understand the engineering behind the obligations the regulation keeps returning to, such as transparency about AI use and human supervision of decisions.
For the regulation's list of high-risk uses (hiring, credit scoring, access to health services and others) and an approval-gate lesson, read our Spanish page curso de agentes de IA para profesionales. Companies thinking about AI beyond agents may also find curso de IA para empresas useful; both describe the same live classes, taught in English.
English at work
Less than many people expect. EF's English Proficiency Index 2025 scores Peruvians aged 26 to 30 at 559, in the "high" band; those aged 31 to 40 score 537 and those over 41 score 509, both "moderate".
Agent platforms, their documentation and their error messages are in English first, and many agents are configured with written instructions. A professional who can write a clear English instruction, and read an English error, works faster with every tool.
EF tests adults who choose to sit its online test, so its scores describe that group rather than every worker in Peru. Still, they suggest many professionals already have the English a live technical class needs.
Mentors speak at a steady pace, check understanding often and write key terms in the chat. Learners describe their own processes in English, which doubles as practice for international meetings.
Connectivity is rarely the issue: INEI found that 95.2% of Peruvians aged 25 to 40 used the internet in the second quarter of 2026, and 86.0% of those aged 41 to 59. We do not teach in Spanish, but one-to-one classes let each learner set the pace.
Judgement first
Not to start. Low-code platforms let you build a working agent without writing code. But you do need to think like an engineer about what happens when things go wrong, and that is what we teach.
Before building, write the steps a person follows today, the systems they touch, and the decisions they make. Most failed agents automate a process nobody had written down.
For every step, ask: does it change something? Can it be undone? Who notices if it goes wrong? Those answers decide where the agent must stop and ask.
Slow systems, missing data, a request in an unexpected format. An agent that has only been tested on good days has not been tested.
Adults who want to add coding to that judgement can take our Python routes, and our Spanish page on cursos de programación para adultos describes how working adults in Peru fit coding into a week. Claude-based tools are used only with learners aged 18 and over, as Anthropic's terms require.
Progress
Our own milestones for professionals, used to set goals and report progress. Not a certification.
| Milestone | You can | Evidence |
|---|---|---|
| 1. Process map | Describe a work process step by step, with its systems | A one-page map |
| 2. First agent | Build an agent that reads and answers | A working demo on your platform |
| 3. Side effects | Separate read-only tools from tools that change things | A tool table with risks |
| 4. Safe retries | Add timeouts, backoff and idempotency keys | The retry lesson, on your process |
| 5. Hand-off | Stop and pass a case to a person with a clear note | Three tested failure cases |
| 6. Present | Explain the agent's limits to a manager in English | A short recorded walkthrough |
A timeout is a question, not an answer. Find out what happened before you try again.
Every agent needs a sentence that begins: "I stopped because".
Ask to see milestone 5 before any agent touches a client.
Progress notes list processes mapped, agents built and failure cases tested.
More courses
Each is live and in English, with a single mentor from start to finish.
AAP / USE / 01
ChatGPT, Claude and automation, used with judgement.
Open the syllabusAAP / USE / 02
Excel, Sheets and Power BI with AI help.
Open the syllabusAAP / USE / 03
Professional graphics with AI features.
Open the syllabusAAP / BUILD / 01
Coded agents with retrieval and evaluation.
Open the syllabusAAP / BUILD / 02
AI coding agents for people who build software.
Open the syllabusAAP / BUILD / 03
The theory behind planning and agents.
Open the syllabusAAP / BASE / 01
From zero to advanced, at an adult pace.
Open the syllabusAAP / BASE / 02
SQL, Python and BI for business data.
Open the syllabusAAP / BASE / 03
The databases agents read from and write to.
Open the syllabusHow classes run
Any day, any time, flexible: choose a batch, or for a very personalised plan choose 1-on-1 at the hour that suits your family.
Before work
An early start in Lima is evening in India, an easy hour to teach.
After work
Evenings in Lima are early morning in India; we can arrange them.
One to one
Built around your role, your platform and your diary.
Projects use processes from your work, anonymised.
Each agent is tested on bad days, not just good ones.
Copilot Studio, Gemini Enterprise or Python.
Clear, steady speech; terms written in the chat.
The same person from first map to final review.
This course is for learners aged 18 and over.
Student work
Projects from learners on our longer courses; see the student labs page for more.

AI and ML
An AI nutrition coach that reads what you eat and works you toward a target.

AI and ML
A chatbot that answers mathematics and programming questions, built and deployed by a student.

AI and ML
An assistant that helps a young person recognise unsafe situations online.

Web app
A weather forecasting site with live conditions for any location.
Fees
Payment is in US dollars by card, and a Peruvian card issuer converts each charge to soles at its own daily rate. The soles figure on your statement may therefore shift a little from month to month.
Free first class
Free
no card required
Group batch
$100 USD
a month, eight live classes
One to one
$150 USD
a month, eight live classes
What families say
Reviews from our learners and families on Google, as written.
★★★★★
"The one step solution for my son. Modern Age Coders make learning coding so simple that kids love it. The teachers explain complex concepts clearly with practical exercises and interactive content."
Ria Mukherjee
Parent
★★★★★
"Modern Age Coders has been a game-changer for me. I struggled to grasp IT concepts and coding before joining, but their classes transformed everything. I can now confidently write complex programs with ease."
Samriddha Mondal
Student
★★★★★
"One of the most wonderful education centres out there. Education is not limited to school syllabus but focuses on skill development."
Vansh Agarwal
Student
★★★★★
"My child Dhairya is really enjoying the Modern Age Coders classes. This is his first online class and he eagerly looks forward to it. I can already see his improvement, and the teachers are very cooperative."
Sonam Oswal
Parent of Dhairya
★★★★★
"Modern Age Coders have wonderful teachers who teach in a clear, easy and practical way. The teacher boosts students' confidence and inspires them to learn without hesitation."
Sonu Goyal
Parent
★★★★★
"I highly recommend this computer coding class! The teachers are incredibly knowledgeable and passionate about coding."
Ritu Kedia
Parent
Professionals in Peru ask
No. Every class is taught live online in English, and there is no Spanish-language version. Mentors keep a steady pace, check understanding often, adapt to your accent and write key terms in the chat, so professionals with working English follow comfortably.
No, for the Copilot Studio and Gemini Enterprise courses, which are low-code. If you want to build agents in Python, our Python Automation course or Python Masterclass gets you there. Either way, we teach the engineering judgement behind reliable agents.
The one your company uses. Copilot Studio suits Microsoft 365 organisations, Gemini Enterprise suits Google Workspace and Google Cloud, and Python suits teams that want full control. The free first class helps you decide.
Timeouts on every tool call, safe retries with backoff and jitter, idempotency keys so a repeat does not repeat an action, a limit on attempts, and a clear hand-off to a person. The lesson on this page shows all of them.
Any day, any time, flexible: choose a batch, or for a very personalised plan choose 1-on-1 at the hour that suits your family. India is ahead of Peru by 10 hours 30 minutes on every day of the year, so 07:30 in Lima is 18:00 in India.
DS 115-2025-PCM sets phased periods for private companies by sector and size, from one to four years after September 2025. Whether and how it applies to your company is a question for a lawyer; we quote it as context and give no legal advice.
Yes, and we recommend it. In one-to-one classes, projects use an anonymised version of a real process from your job. Please never share client personal data in class.
Yes. The professionals' agents courses are for learners aged 18 and over, and Claude-based tools are used only with adults, as Anthropic's terms require. Teenagers have separate routes.
You receive our own certificate of completion. It is not a Microsoft or Google certification and not a Peruvian qualification; we are not connected with Microsoft, Google, AWS, OpenAI or any Peruvian authority.
A laptop or desktop, a stable connection and a headset. For platform courses you need access to the platform, through your employer or a trial account; we help you set it up in the first class.
Related
More from our site.
Everything we teach in Peru.
Age rules for every AI tool.
Judgement before tools.
Projects by our learners.
The live class format.
Every syllabus in one place.
Part of the guide: Every Peru page
More guides: Coding and AI classes in Peru · En español: clases de programación en inglés · Start here: English-medium coding in Lima · By age and stage: AI and coding for professionals · Exams, olympiads and school holidays: IB computer science tutor · Guides to choosing: An English-taught option to Crack The Code · Regions of Peru: Amazonas region · Cities: Arequipa: coding for every age · Lima and Callao districts: Programación para niños en Ancón
Start here
Bring one process from your work. In the trial, you and a mentor map it, mark the steps that change things, and sketch where an agent could help and where it must stop. You leave with a recommended platform and course.
If you would like to read first, how we teach explains the format and the course catalogue lists every syllabus.
WhatsApp us · +91 91233 66161 · connect@modernagecoders.com
Classes are live video sessions in English with mentors based in India. Message us on WhatsApp in English for a fast reply.
A priority demo is a full class of about 45 to 60 minutes. After the class the mentor sends a written skill report and a personal learning roadmap on WhatsApp. If you enrol, the demo fee is adjusted against your first month's fee. Share your demo Order ID with us on WhatsApp and we take it off.
Book a priority demo