AI book platform
Personalised kids' books with AI face swapping
Laravel 13 · Fabric.js · Segmind AI
I'm Kafi, a senior freelance developer. My AI agent development services build custom agents that take a goal, plan the steps and use your tools to finish the job, with limited permissions, human approval on risky actions and a log you can read afterwards.
Starting from $10/hr
A narrow goal and a short list of allowed tools
Permissions limited to what the job needs
Approval before money, deletion or customer messages
A full log of every step, input and output
AI AGENT DEVELOPMENT
An AI agent is software that uses a language model to decide which steps to take, then calls tools such as your database, email, calendar or payment system to carry them out, until a goal is reached or it needs help. A chatbot answers questions; an agent does work.
That freedom is what makes agents useful and risky. A good agent is designed around limits: a narrow goal, a short list of tools, the least access that works, approval for anything irreversible and a log of every action.
I build agents in layers. The tools are plain, tested code. The model decides the order. The guardrails live outside the model, so a bad answer cannot become a bad action.
| Custom agent | General agent tool | |
|---|---|---|
| Scope | One job, defined tools | Tries to do everything |
| Permissions | Least access, scoped per tool | Broad account access |
| Approval | Built in for risky actions | Optional or missing |
| Reliability | Tested on your real cases | Unpredictable on edge cases |
| Audit | Full log you can review | Limited visibility |
WHAT IS AN AI AGENT?
An AI agent is software where a language model decides what to do next and uses tools, such as your database, calendar or email, to get a task done. A general guide:
| Chatbot | Automated workflow | AI agent | |
|---|---|---|---|
| Who decides the steps | The user, one message at a time | You, in advance | The model, within limits you set |
| Uses your tools | Rarely | Yes, fixed steps | Yes, chosen as needed |
| Predictability | High | Highest | Lower: needs limits, approvals and logs |
| Best for | Answering questions | Repeatable processes | Tasks where the path varies |
| More | AI chatbots | AI automation | — |
HOW AN AGENT RUNS
An agent is software that plans steps and calls your tools to reach a goal. What makes it safe is what it is allowed to touch and who approves the risky steps.
Example run · your tools, permissions and approval rules are set per project
WHAT I BUILD
Agents work best on defined jobs with clear tools. These are the kinds I build.
An agent designed around one of your processes, with the tools, rules and approvals that process needs.
Custom · Tools
Collect, compare and summarise information from sources you approve, with citations.
Research · Citations
Research a lead, draft personal outreach and update the CRM, with you approving what is sent.
CRM · Outreach
Read a ticket, look up the order, check policy and prepare the resolution for approval.
Tickets · Orders
Reconcile invoices, chase payments, update records and flag exceptions.
Finance · Ops
Agents that branch, retry and wait for people inside a longer process.
Workflows · Retries
Your systems exposed to an agent through the Model Context Protocol, with scoped access.
MCP · APIs
An in-app agent that can act for the signed-in user, within their permissions.
In-app · Auth
Agents that talk on the phone and use the same tools, with transfer to a person.
Voice · Phone
A set of real test cases, scored on every change, so you can see if the agent got better or worse.
Evals · Tests
Spending limits, allow-lists, approval queues and kill switches built in from day one.
Safety · Approvals
Logs, traces, alerts and per-run cost tracking so nothing runs away.
Logs · Costs
Describe what you want to build. I'll reply within a few hours and tell you honestly what it would take.
TYPES OF AI AGENTS
Textbooks list several agent types. In business projects they usually mean this:
| Textbook type | What it means | Where it shows up |
|---|---|---|
| Simple reflex | Reacts to a trigger with a fixed rule | Alerts and routing rules |
| Model-based | Keeps track of state to decide | Order or ticket status handling |
| Goal-based | Plans steps toward a stated goal | Booking, research, data gathering |
| Utility-based | Picks the best option by a score | Prioritising leads or tasks |
| Learning | Improves from feedback | Systems with review loops |
| Multi-agent | Several agents with separate roles | Larger pipelines with checks between steps |
HOW IT WORKS
Start narrow, prove it, then widen.
A job with clear success criteria and a real cost today.
The exact actions the agent may take, each with its own limits.
Plain, tested code for every action, so the model never touches your systems directly.
Real examples with expected outcomes, including hard and hostile ones.
The agent proposes, you approve, until the results earn trust.
Logs, cost tracking and alerts, then loosen approvals only where it has proven reliable.
GUARDRAILS
An allow-list of tools, per-action limits and human approval for anything costly or irreversible.
Step and spend ceilings per run, with an automatic stop and an alert.
The agent sees only the records its task needs, with role-based access. Regulated data goes through your own compliance review.
Every step, tool call and result is logged so you can replay what happened.
Untrusted text is treated as data, never as instructions, and risky actions need confirmation.
Start with one narrow task, measure it, then widen.
TOOLS
Reliable agents are mostly ordinary engineering.
PROOF
Real products in production, described without client names. Each one started with a conversation like the one you're about to have.
Personalised kids' books with AI face swapping
Laravel 13 · Fabric.js · Segmind AI
Member, plot and transaction management in 2 weeks
Laravel
Live classes built in (Zoom)
Laravel · Zoom SDK · Docker
HOW TO CHOOSE
Whether you hire a freelancer, a studio or a large consultancy, these questions separate demos from working systems.
A good partner narrows the first agent to one job with a measurable result.
Ask about tool allow-lists, spend caps, approvals and logging, not just the model name.
Look for a set of real example cases run before launch and after each change.
Repository, prompts and accounts in your name, with documentation.
The value is in your CRM, helpdesk and database, so integration skill matters.
Models and APIs change; agree who monitors and updates the agent.
COST & TIMELINE
It depends on the number of tools the agent uses and how much testing the job needs. A single-job agent with three or four tools is moderate; an agent that touches money or customer data needs more safeguards and testing. You get a written estimate after a short call, and hourly work starts at $10 when you hire me directly.
Running costs are model usage per run, billed by the provider to your account. I log cost per run, so you can see what each task costs before you scale it.
Each tool is built, tested and permission-scoped separately.
Money, deletion or customer messages need approvals and extra tests.
More edge cases mean more evaluation work, and a more reliable agent.
Sensitive data needs masking and tighter permissions.
Drives model cost and the need for queues and monitoring.
WHO IT'S FOR
Where agents earn their keep.
PRICING & PLATFORMS
Work with me directly for the best rate, or through Upwork or Fiverr if you want the platform's payment protection.
From $10/hr
Upwork
Fiverr
Long-term work starts at $10 per hour. Every project gets a written estimate before it starts, so there are no surprises.
FAQ · 12 QUESTIONS
Still unsure? Book a free call and ask me directly. I'll tell you honestly if I'm the right fit.
It is building software that uses a language model to plan steps and call tools, such as your database, email or calendar, to reach a goal. The developer defines the tools, limits and approvals that keep it safe.
It is a program where a language model decides the next action, the code executes it through defined tools, and the result feeds back into the model until the task is complete or a person is needed.
It depends on the number of tools, the risk of the actions and how much testing the job needs. I give a written estimate after a short call, and hourly work starts at $10 when you hire me directly.
A focused agent with a few tools can be working in shadow mode in a few weeks. Agents that act on money or customer data take longer because of testing, and you get a timeline in writing.
Pick one job, list the allowed tools, build and test each tool, write real test cases, run the agent in shadow mode with approval, then go live with logging and cost limits.
Ask how they limit permissions, how they test, whether risky actions need approval, what is logged and who owns the code. Look for narrow, proven examples rather than promises of a general agent.
The Model Context Protocol is an open standard for exposing tools and data to AI models. I build MCP servers so an agent can use your systems through a scoped, documented interface.
For a defined business process, usually yes: you control the tools, permissions and approvals, and it is tested on your real cases. General agent tools are fine for experiments.
Only with limits. Mine get the least access that works, cannot take irreversible actions without approval, and log every step, so you can review what happened and stop it at any time.
Step and spend limits per run, allow-listed tools, timeouts and alerts. If a run exceeds a limit, it stops and reports instead of continuing.
A chatbot mainly answers questions in conversation. An agent also takes actions across your tools, such as updating a record or preparing an invoice, so it needs stronger permissions and approvals.
Yes. The code, prompts, tools and accounts are yours, documented so another developer can maintain them.
Tell me what you're building and where you're stuck. You'll get an honest plan, timeline and estimate, even if I'm not the right fit.