Loading Studio Assets...
AI agents are the most transformative development in software since the smartphone — and also the most in-demand skill in the tech job market. Senior AI agent developers are commanding Rs 50-80 lakh packages at Indian unicorns, and Rs 1 crore+ at global companies.
Here is what most tutorials miss: building a basic AI agent is not as hard as it looks. The concepts are clear, the tools are mature, and with $5 in API credits you can have something working today.
This guide is specifically written for Indian developers — with use cases relevant to the Indian market and infrastructure assumptions that match Indian development environments.
A chatbot receives a message, generates a response, and stops.
An AI agent receives a goal, decides what actions to take, uses tools to take those actions, evaluates results, takes more actions, and reports when done.
The key difference is autonomy and tool use. An agent can search the web, read files, call APIs, execute code, interact with databases, and use other AI models as sub-agents — in sequences, without human intervention at each step.
1. The brain (LLM) — The AI model that reasons and decides what to do next. GPT-4o, Claude, or Gemini can serve as the brain.
2. The tools — Python functions that the agent can call to interact with the world.
3. The memory — What the agent knows about past actions and their results (stored in the messages list).
4. The orchestrator — The code loop that calls the LLM, executes tool results, and manages state.
A Research Agent that:
This same architecture adapts to competitor intelligence, lead research, market analysis, and news monitoring.
Create a project folder and install dependencies:
pip install openai langchain langchain-openai duckduckgo-search requests beautifulsoup4 python-dotenv
Create a .env file with your OPENAI_API_KEY.
For Indian developers: HDFC, ICICI, and SBI Visa credit cards work for OpenAI API. If your card is declined, a Wise virtual card (free to create) is the most reliable alternative.
Tools are Python functions with descriptive docstrings. The agent reads the docstring to understand what each tool does and when to use it.
Your research agent needs three tools:
search_web(query) — Searches DuckDuckGo using the duckduckgo-search library and returns titles, URLs, and snippets for up to 5 results.
read_webpage(url) — Fetches a URL using requests and BeautifulSoup, strips navigation and scripts, and returns cleaned text limited to 3,000 characters to avoid exceeding the context window.
save_report(filename, content) — Writes the final markdown report to a local file.
The docstrings are critical. Write them clearly — they are the agent's documentation for how and when to use each capability.
This is the core of the pattern:
def run_agent(research_question, max_iterations=10):
messages = [
{"role": "system", "content": "You are a thorough research agent..."},
{"role": "user", "content": f"Research this topic: {research_question}"}
]
for iteration in range(max_iterations):
response = openai_client.chat.completions.create(
model="gpt-4o",
messages=messages,
tools=TOOLS,
tool_choice="auto"
)
message = response.choices[0].message
messages.append(message)
if message.tool_calls:
for tool_call in message.tool_calls:
tool_name = tool_call.function.name
tool_args = json.loads(tool_call.function.arguments)
result = TOOL_MAP[tool_name](**tool_args)
messages.append({
"role": "tool",
"tool_call_id": tool_call.id,
"content": str(result)
})
else:
return message.content # Agent finished
The LLM decides at every step: search the web, read a page, or save the report. You did not hard-code the sequence — the agent reasoned it out from the available tools and the goal.
That is what makes it an agent rather than a script.
OpenAI's tool calling requires you to describe each tool in a JSON schema. For each tool, provide:
name: Must match your Python function namedescription: What the tool does (the LLM reads this to decide when to call it)parameters: JSON schema for the function's argumentsThis schema is what gets sent to the model alongside your messages. The model returns a structured tool call object — your code executes the function and appends the result back to the messages.
python agent.py
Expect output like:
Starting Research Agent
Question: What is the state of the Indian EV market in 2026?
Iteration 1 - Using tool: search_web
Iteration 2 - Using tool: search_web
Iteration 3 - Using tool: read_webpage
Iteration 4 - Using tool: read_webpage
Iteration 5 - Using tool: save_report
Agent completed! Report saved as india_ev_market_2026.md
Total API cost for a typical research task: $0.05 to $0.15.
Notice: you never told it to search first, then read pages, then save. It figured out that sequence from the tools available and the goal. That emergent planning is the core capability that makes agents powerful.
Once comfortable with this pattern, here is how to grow it:
Add more tools — Email sending via SendGrid, database queries via SQLAlchemy, Slack messaging, Google Calendar — any API becomes a tool with a few lines of code.
Persistent memory — Use a vector database like Chroma or Pinecone to give the agent memory across sessions. It can remember previous research and build on it.
Try the Claude API — Anthropic's API works with nearly identical code structure but often produces better reasoning for complex research tasks. Switch by changing the client and model name.
Use LangChain — The LangChain framework provides pre-built agent implementations, a library of ready-made tools, and observability dashboards. It reduces boilerplate significantly.
Multi-agent systems — One orchestrator agent can spawn sub-agents for different tasks. A research agent calls a writing agent calls a fact-checker agent. This is how enterprise AI systems are built.
Here are agent ideas with genuine commercial potential in India:
GST Reconciliation Agent — Reads invoice PDFs, matches entries against GSTR-2A data, flags discrepancies, generates a reconciliation report. Massive pain point for every CA firm and SME in India.
WhatsApp Business Agent — Handles inbound customer queries via the WhatsApp Business API, qualifies leads, books appointments, escalates to human agents when needed. Reaches India's 500M+ WhatsApp users.
Tender Monitoring Agent — Watches GeM, CPPP, and state procurement portals for relevant government tenders, sends daily alerts, drafts initial compliance checklists.
Social Media Monitor — Tracks brand mentions across Twitter/X, LinkedIn, and YouTube. Summarizes daily sentiment, flags negative threads that need response.
Job Board Aggregator — Monitors Naukri, LinkedIn, Indeed, and AngelList for roles matching custom criteria. Sends a filtered daily digest. The search process that takes 30 minutes manually takes 30 seconds with an agent.
This tutorial gives you the foundation for one of tech's highest-demand roles right now:
The skills from this tutorial — tool use, the agent loop pattern, state management, LLM integration — are the core building blocks for all of these roles.
Over-engineering too early — Build the simplest version that works first. Add complexity when you hit real limitations, not before.
Ignoring cost — Every LLM call costs money. Add token counting and cost logging from the start. Without monitoring, an agent that loops unexpectedly can burn through API credits fast.
No max_iterations limit — Always set a maximum iteration count. Without it, a confused agent can loop indefinitely.
Trusting tool outputs blindly — Tool results can contain errors, empty responses, or irrelevant content. Build error handling and teach your system prompt how the agent should respond to failure cases.
Skipping evaluation — How do you know if your agent is actually good? Build a small test suite: give it 10 known research questions and check whether the outputs are accurate and well-structured.
The architecture you just learned powers AI systems at Salesforce, HubSpot, Razorpay, and Zepto. It is not magic — it is a clear pattern that scales from a 50-line tutorial to a million-dollar enterprise system.
The distance between this tutorial and production-ready AI agents is shorter than you think. The main ingredients beyond this foundation are: better prompts, more tools, error handling, and evaluation.
Start with this research agent. Adapt it to one real problem you or someone you know faces. Then go build something India actually needs.
Turn your AI skills into products that matter. Brandomize covers the tools, trends, and strategies helping Indian developers and entrepreneurs build in the AI era.
We help founders, brands, and local businesses turn modern tech into measurable revenue and standout brand identity.
AWS signed a multiyear deal with vibe-coding startup Superblocks to run AI app-building inside customers' own private clouds — so 'data never leaves.' Here's why it matters.
Engineers no longer type most code by hand — they describe intent and agents do the work. Here's how the three leading agentic coding tools differ in 2026.