Agent development
AI agent development
what gets built, and how it is tested
Written for the reader who will ask what runs where. An Agent is a model that works out what was meant, code that makes the result correct, and a log of everything it did.
An Agent is three things. A model that works out what was asked, code that makes sure the result is correct, and a record of every action it took. Build it the other way round and you get a menu.
I build custom AI agents for Indian businesses, on my own or alongside a client's IT team. This page is the technical version: what sits inside an Agent, which systems it reads and writes, how it is tested before anyone relies on it, and what is handed over at the end. If you want the plain version first, read what an Agent does in an office.
The model decides, the code guarantees
The model reads the message and works out the intent. Ordinary code does the arithmetic, the lookups and the writing, so the same request gives the same result every time it is asked.
It asks before it changes anything
Any action that writes a record or sends a message outside the company is confirmed by a named person first. The confirmation and the action both land in the log.
Tested against your real records
I run it over your own history rather than sample data, and compare what it produced with what your team actually did. The failures get fixed before anybody depends on it.
One Agent, several surfaces
The same Agent can answer in WhatsApp, in a team chat on Discord, in email and in a sheet. The rules live in one place, so there is nothing to keep in step.
It can run on your own machine
An Agent can call a model over the internet, or run on an open model on a computer in your office. That is a build decision, and I will tell you which one your case needs.
Handover is a deliverable
The code, the accounts, the configuration, a written runbook and the list of what to check when something breaks. All of it in your name, and all of it readable by your own developer.
What an Agent is made of
There is no keyword matching anywhere in an Agent I build, and no written replies waiting to be picked. The model is what understands a message typed by a person in a hurry, in English or in Hinglish. Code is what decides whether the resulting action is allowed, does the sums, and writes the record. If the model is unsure, the Agent asks the person rather than guessing.
Every Agent I build has the same six parts. The interesting work is in the contracts between them, because that is what stops a confident wrong answer from reaching a real record.
- An entry point: a WhatsApp number, a chat channel, an email address, a form, or a scheduled run
- A model that reads the request and works out what is being asked for
- Tools the model may call, each one a small piece of code with a fixed input and output
- The records: your sheets, drive, calendar, database or accounts software
- A confirmation step for anything that changes a record or leaves the building
- A log of every request, every tool call and every result
What it connects to
The Agent reads and writes where your work already sits. WhatsApp on your own business number, a team chat on Discord, email, Google Sheets, Google Drive, Google Calendar and Notion. Accounts software including Tally can be connected as part of a build, as can an existing database, a file server, or a fingerprint attendance device on your network. I do not sell voice or calling Agents, so everything an Agent of mine does happens in writing.
Each connection is scoped. The Agent gets read access to the folders, groups and sheets you name and nothing beyond them, write access only where you have asked for it, and that list is visible and editable afterwards. The WhatsApp Agent page covers the one surface most Indian businesses start with.
How it is tested before anyone trusts it
First it is replayed over your own history. I take the last stretch of real messages, reports or enquiries, run the Agent over them, and compare its output against what your team actually produced. That is where the wrong assumptions show up, and they always do.
Then one team uses it on live work while I watch the log. Mistakes get fixed, confirmation steps get tightened where people were clicking through them, and the rules get written down as they settle. The last test is boring and it matters: a power cut, a restart, a day with no internet, and a week with nobody babysitting it.
Where the Agent runs, and the training question
Two choices. The Agent can call a model over the internet, which is simpler and cheaper to start. Or the whole thing can run on a machine in your own office using an open model that never contacts an outside service, which is what businesses choose when the records cannot leave the building.
Training a model on your own material, which is called fine-tuning, is a third option and a narrower one. It is worth doing for one specific internal job on hardware you own, and most briefs that ask for it actually need retrieval instead: the model reads your documents at the moment of the question. I run fine-tuning on my own machine today. No client has bought fine-tuning from me yet. Treat it as something I can build rather than something with a case study. Fine-tuning or retrieval is the honest comparison.
How a build is quoted
Per project, after a scoping pass, and the quote names every surface, every system being connected, who confirms what, and the handover. The schedule comes with it. There is no monthly software subscription on what I build, and nothing is charged for each person who uses it.
What moves the number: how many jobs the Agent has to do, how many systems it touches, how clean those records are today, whether it runs on an outside model or on hardware you buy, and whether you want me on call afterwards. Care is quoted on its own and can be stopped without the Agent stopping.
How it runs
- 01
Scope
We write down what the Agent must do, which systems it touches, who confirms what, and what a correct result looks like for each job.
- 02
Build against your records
I build it and point it at copies of your real data, so the first version meets the mess that actually exists rather than a clean example.
- 03
Run it with one team
One team uses it on live work while I read the log. Mistakes get fixed and the rules get written down as they settle.
- 04
Handover
Code, accounts, configuration, runbook and training. Your developer or your IT vendor can take it from there without calling me.
Built, not described
Questions people ask
Who builds AI agents for Indian businesses?
Most of the firms selling agent development in India are software agencies running a team of developers. I am one person who does the scoping, the build and the handover, which suits a business that wants the builder in the room. For a programme needing twenty people at once, I am the wrong call and will say so.
How much does it cost to build an AI agent?
Quoted per project after scoping, never per hour, and the quote itemises every surface and every system being connected. The things that move it are the number of jobs, the number of connections, how clean your records are, and whether it runs on an outside model or hardware you own. The schedule comes with the quote.
What do you build Agents on?
For work that may use an outside model, I build on Claude. Where the records have to stay in your building, the same Agent runs on an open model on your own machine. The surrounding code is ordinary code in your repository, so there is no platform licence sitting underneath your system.
Can the Agent write into Tally or our existing database?
Yes, as part of a build. Accounts software including Tally, an existing database and a file server can all be written to, with the write path scoped and a confirmation step in front of anything that changes a financial record. I have not built a Tally connection for a client yet. Ask me how I would do it rather than for an example of one I have done.
How do you stop an Agent from doing something wrong?
Three ways. The model never performs an action directly: it calls a tool, and the tool's code decides whether the action is allowed. Anything that changes a record or leaves the company waits for a person to confirm. And everything is logged with the record it touched, so a bad result can be traced instead of argued about.
Can our own developers maintain it afterwards?
That is the point of the handover. The repository, the configuration, the accounts and a written runbook are yours, in plain code with no hidden platform in the middle. I will walk your developer or your IT vendor through it, and after that they do not need me to change a rule.
Do you fine-tune a model on our data?
I do, on hardware you own, for one narrow internal job where it genuinely helps. Most requests for it are better served by retrieval, where the model reads your documents when the question is asked. No client has bought fine-tuning from me yet, and I would rather say that than imply otherwise.
Can you work alongside our IT team?
Yes, and it makes the handover easier. Your team usually owns the network, the accounts and the existing database, and I build the Agent and the connections into them. Where you have a developer, I would rather they reviewed the code during the build than met it at the end.
Next step