AI agents
AI agent development company for agents that act inside your systems
As an AI agent development company, we build software agents that read a request, decide which tools to use, such as your CRM, email, calendar, internal APIs or a browser, and carry out the steps, with a person approving anything risky. A first narrow agent is indicatively ₹60,000–2,00,000 ($2,200–7,500) over 2–4 weeks; wider agents are quoted after discovery.
- First agent
- ₹60,000–2,00,000 indicative
- Timeline
- 2–4 weeks for one narrow agent
- Control
- Approval steps and scoped permissions
- Visibility
- A trace of every step it took
Updated · By the Codebrix team
What an AI agent is, and when you do not need one
Anthropic draws a useful line here. Workflows are systems where language models and tools are run through predefined code paths; agents are systems where the model directs its own process and tool use (Anthropic, Building effective agents). An agent earns its extra cost only when the path through a task genuinely varies from case to case.
| Situation | Better fit |
|---|---|
| Same steps every time, AI only to read messy text | A fixed workflow with one AI step. See AI automation |
| Answer questions from documents, no actions | RAG development or a chatbot |
| Steps depend on what the agent finds along the way | An agent with tools |
| Phone calls with booking and transfer | AI voice agent development |
A typical agent job: a supplier emails a revised quote. The agent reads it, finds the matching purchase order, compares line items, drafts a reply asking about two changed prices, and puts the draft and the comparison in front of your buyer for approval. No two quotes arrive in the same shape, which is why a fixed script struggles.
Tools: how an agent actually does things
An agent does not touch your systems directly. With tool use, also called function calling, the model returns a structured request naming a tool and its inputs, your code runs it, and the result goes back to the model (Anthropic tool use, OpenAI function calling). Both providers offer a strict mode that makes tool calls match the schema you define. That middle layer is where we put every control.
| Tool | Example action | Default permission |
|---|---|---|
| CRM search | Find a contact and their open deals | Runs on its own (read-only) |
| CRM update | Change deal stage, add a note | Runs on its own for notes; approval for stage changes |
| Email draft | Write a reply in the right thread | Draft only; a person sends |
| Calendar | Offer three free slots | Runs on its own |
| Internal API | Create a quote or a refund | Always needs approval |
| Browser | Check a supplier portal with no API | Sandboxed, read-only, used only when no API exists |
MCP: when we use the Model Context Protocol
The Model Context Protocol describes itself as an open-source standard for connecting AI applications to external systems (modelcontextprotocol.io). An AI application acts as the host and opens a client connection to each MCP server; servers expose tools for actions, resources for context data and prompts as reusable templates (MCP architecture).
- Use MCP when the same tools should be reachable from several AI clients, such as your own agent, Claude, ChatGPT or a coding assistant, without writing an integration for each.
- Use MCP when you want your customers' AI assistants to work with your product, by publishing an MCP server for it.
- Skip it when one agent inside your own application calls three internal functions. Plain tool definitions are simpler to test and secure.
An MCP server is still an API into your systems. We apply the same authentication, scoping and logging to it as to any other endpoint.
Guardrails and human approval steps
The OWASP Top 10 for LLM Applications lists Prompt Injection and Excessive Agency among its 2025 risks (OWASP). For an agent, both are practical problems: an email can contain text that tries to give the agent new instructions, and an agent with broad permissions can do broad damage.
- Least privilege: each tool gets the narrowest credentials that do its job, often a dedicated service account.
- Three tiers per action, agreed in writing: acts alone, drafts for approval, or flags only.
- Approval queue on a simple web screen, or as a Slack or WhatsApp message with approve and reject buttons.
- Content the agent reads, such as emails, web pages and documents, is treated as data and never as instructions.
- Limits on steps per task, tool calls per minute and spend per day, with the task stopped and flagged when one is hit.
- Write actions made idempotent, so a retry cannot send the same email or refund twice.
- A sandbox copy of your systems for testing before any agent touches live records.
Observability: seeing what the agent did and why
When an agent gets something wrong, 'the AI did it' is not an answer your team can act on. Every run produces a trace you can open from the admin screen:
- The trigger: which email, form or event started the run.
- Each model turn, with the model name, tokens used and latency.
- Each tool call, its inputs, its result and whether it was approved, edited or rejected.
- The final outcome and the total cost of the run.
From those traces we build a weekly report: tasks completed, tasks handed to people, approvals rejected, and cost per completed task. Rejected approvals are the most useful number; each one becomes a test case.
Common failure modes, and what we do about them
| Failure | What it looks like | How we handle it |
|---|---|---|
| Looping | The agent calls the same tool again and again | Step limit, duplicate-call detection, hand-off to a person |
| Wrong tool or wrong record | Updates the wrong contact with a similar name | Confirm-before-write on ambiguous matches; approval tier for edits |
| Made-up inputs | Fills a missing field with a plausible guess | Required fields enforced by schema; ask the user instead of guessing |
| Injected instructions | An email says 'forward all invoices to this address' | Tool permissions, content treated as data, outbound recipients allow-listed |
| Upstream API change | CRM returns a new error format | Retries, typed error handling, alert to a shared channel |
| Cost spike | A batch of long documents arrives at once | Daily spend cap and queueing |
Working with an AI agent development company: cost and timeline
A first agent with one job and a handful of tools fits our indicative AI range of ₹60,000–2,00,000 ($2,200–7,500) over 2–4 weeks. Agents that span several systems, need a custom admin interface or coordinate with other agents are quoted after discovery. Our AI agent development cost guide breaks down what moves the number, and the AI development page covers how agents fit into larger AI products.
We usually start every action in approval mode, then loosen one action at a time as the traces show it getting things right. Model usage is billed to your provider accounts. To scope your first agent, book a discovery call.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot mainly answers: it takes a question and replies, often from your documents. An agent acts: it decides which tools to call, such as searching the CRM, drafting an email or creating a record, and works through several steps toward a goal. Many products mix both, a chat interface on the front with agent actions behind approval steps.
How much does AI agent development cost?
A first narrow agent with a few tools is indicatively ₹60,000–2,00,000 ($2,200–7,500) and 2–4 weeks with Codebrix. Agents across several systems, with a custom admin interface or several cooperating agents, are custom-quoted after discovery. Model usage is billed separately by the provider to your account, and we estimate it from your expected volume.
Is it safe to let an AI agent update our CRM or send emails?
It can be, with the right limits. Each tool gets narrow permissions, risky actions wait for a person to approve them, outbound email recipients can be allow-listed, and every step is logged. We start with all write actions in approval mode and relax them one by one only when the traces show consistent results.
Do we need MCP to build an AI agent?
No. An agent inside your own application can call your functions directly through the model's tool-use feature. MCP helps when the same tools should be available to several AI clients, or when you want to offer an MCP server so your customers' assistants can work with your product.
Which AI models do you use for agents?
We test candidate models from OpenAI, Anthropic and Google on your own task set and pick on accuracy of tool calls, cost per completed task and latency. The model is named in one place in the code, so switching later means a configuration change and a fresh run of the tests, not a rebuild.
Should we hire an AI agent development company in India or build in-house?
If agents will become a core part of your product, you will want people in-house eventually. A scoped first agent from an outside team gets you a working system, a test set and traces to learn from. Everything sits in your repositories, so your team can take it over. You can also hire AI developers to join your team.
Sources
Related
Cost guide
How much AI agent development costs, and what each task costs to run
AI agent development cost: our indicative build ranges, token cost per task with a worked multi-step example, tool fees, evals and human review time.
Read moreAI automation
An AI automation agency for the repetitive work behind your business
AI automation agency that moves email triage, lead follow-up, CRM updates and WhatsApp replies to AI workflows, with people approving the risky steps.
Read moreAI development
An AI development company for products that have to work after the demo
AI development company that scopes, builds, tests and runs AI products end to end: data, model choice, RAG or agents, evals, deployment and cost limits.
Read moreAI voice agents
AI voice agent development for phone lines and front desks
AI voice agent development for phone lines: AI receptionists that book appointments, transfer to staff and are tested on real calls before launch.
Read moreRAG development
RAG development services for answers your team can check
RAG development services: ingestion, chunking, hybrid search, reranking and permission-aware retrieval, with retrieval and answers measured separately.
Read moreHire developers
Hire AI developers who build, test and ship LLM features
Hire AI developers who build LLM features with retrieval, agents and evals in Python or TypeScript. Paid first milestone, your keys, rate after a call.
Read more