AI agent development — software that acts, and knows when to stop

Most AI agent development sold in India today is a demonstration. A model is connected to a tool, it performs a task impressively once, and the recording ends before anyone asks what happens on the four-hundredth run when the input is malformed and the action cannot be undone. Building an agent is easy. Building one you can leave running is a different discipline, and almost all of the work is in the second.

What an agent is

Ordinary software follows a path a programmer wrote. An agent is given a goal and works out the path itself: which tool to call, in what order, with what arguments, and when it is finished. The tools are the interesting part — your database, your email, your spreadsheets, your document store. Without tools an agent is a chat window. With them, it is a colleague who is fast, tireless, and occasionally confidently wrong.

The only design decision that matters

Before a line of code is written, every action the agent could take is sorted into two piles.

• Reversible. Drafting text. Tagging a record. Scoring a lead. Summarising a document. Sorting a queue. If the agent gets these wrong, someone fixes it in a minute. The agent does these alone.
• Irreversible. Sending money. Emailing a customer. Publishing under your name. Deleting anything. Committing to a price. If the agent gets these wrong, you are apologising, refunding, or in a dispute. A human approves every one of them.

This sorting is the architecture. It cannot be retrofitted, because an agent designed to act autonomously has no place to put the approval step and no record of what it was about to do. If a vendor describes an agent as "fully autonomous" across actions in the second pile, they are describing a liability and calling it a feature.

Where the cost actually is

Two numbers, always quoted separately, and you should demand the same from anyone you speak to.

The build is a fixed price, agreed after a call. It is driven almost entirely by integration — how many systems the agent must read from and write to, and how badly documented they are. The reasoning itself is rarely the hard part. An agent that touches one spreadsheet is cheap; an agent that touches a CRM, an email provider and an industry tool with no public API is not.

The running cost is the AI provider's metered usage. It is ongoing, it scales with how much the agent does, and it is billed to your own provider account rather than marked up through the studio. There is a direct reason for that policy, described below.

What has been built, including what stopped

CyberSathi — an agent with an editor

CyberSathi.in ran a scheduled pipeline that pulled cybercrime news and video from several sources, identified what was trending, and drafted articles. The drafts were not written in a generic model voice: a style guide was extracted from the client's own earlier blog posts, so the output carried his phrasing rather than the flat register every language model defaults to. Nothing published itself. Every draft went to an admin panel and a person approved or killed it.

It is switched off. The provider's usage was billed to a personal account on a side project; when the credits stopped, the pipeline stopped. The code is intact and would resume on funding. This is the failure that produced the billing policy above — an automation whose running cost sits on someone else's card is an automation with a hidden expiry date.

SkillCircle — the part that actually breaks

A job-drive portal with registration, timed test delivery and results. Over 500 students registered; more than 50 sat the test at the same moment. It shipped with tab-switch detection, window-minimise alerts, locked navigation and timed sessions. No language model anywhere in it — and it is the most instructive project on this page, because concurrency, state and abuse are what break real systems. An agent that works in a demo has not met any of those.

The studio's own agents

The content workflow behind this studio runs on a self-hosted n8n instance on a Mac mini, with Notion as the database and the approval queue, publishing on a schedule to LinkedIn and a blog. Scheduled trigger, fetch, generate, human approves, publish. The design recommended to clients is the one running here, which is the only reason it can be described honestly.

What breaks, and how you find out

• Models get deprecated. The one you built on will be retired. Budget for a migration you did not ask for.
• Output drifts. The same prompt on a newer model version produces different text. Without logs of inputs and outputs you will not know when it changed, only that something feels off.
• Costs scale with success. An agent that processes ten documents a day costs nothing. The same agent at a thousand documents a day is a line item. Model the second case before building.
• The process was never written down. The most common failure is not technical. The rules lived in one experienced person's head, nobody wrote them down, and the agent faithfully automated a misunderstanding.

When not to build an agent

When the task runs twice a month. When an off-the-shelf tool already fits and you are building custom out of preference. When the decision needs judgement the business cannot articulate. When a database and a form would do — which is more often than anyone selling AI would like to admit, and it is the answer you will get on the call if it is the true one.

Frequently asked questions

What is an AI agent, in plain terms?

A program that takes a goal, decides which steps to run, calls the tools it needs — your database, an email API, a spreadsheet, a document store — and produces a result. The difference from ordinary software is that the sequence of steps is decided at run time by a language model rather than fixed in advance by a programmer. That flexibility is the whole value, and it is also the whole risk.

How is this different from a chatbot?

A chatbot answers. An agent acts. A chatbot that has been given the ability to issue a refund, update a CRM record or send an email is no longer a chatbot, and it needs to be engineered as though it will occasionally do the wrong one of those things — because it will.

How do you stop an agent doing something expensive and wrong?

By deciding, before writing any code, which actions are reversible. Reversible actions — drafting, tagging, sorting, summarising — the agent performs alone. Irreversible ones — moving money, contacting a customer, publishing under your name, deleting anything — stop at a human approval step. This is a design decision, not a feature, and it cannot be added afterwards.

What does AI agent development cost in India?

The build is a fixed price, quoted after a call, and driven mostly by how many systems the agent has to connect to rather than by how sophisticated the reasoning is. Integration is the expensive part. The AI provider's usage is metered, ongoing, and billed to your own account rather than marked up through the studio. Any vendor who quotes a single blended number is hiding which half grows over time.

Whose accounts and code is it?

Yours. Source code in a repository in your business's name, workflow definitions you can export, and API keys on your own provider account. An agent you do not own is a process your business depends on running on a machine you cannot reach — a worse position than not having automated at all.

What happens when the model provider changes something?

It breaks, eventually, and you should plan for it. Models are deprecated, prices change, and outputs drift between versions. Agents built here log their inputs and outputs so that when behaviour changes you can see what changed rather than guess. Anyone who has not mentioned this to you has not run an agent in production for long.

Related

• AI automation agency in India — the wider picture, and how to evaluate any vendor including this one.
• AI automation agency in Noida — if you would rather work with someone an hour away.
• All services — websites, portals, SEO and the rest.