AI agent development, scoped to what you allow and logged line by line

AI agent development is building software that takes a goal, decides how to reach it, and then acts inside your real systems: your CRM, your inbox, your database. Kaishift builds those agents from Casablanca and runs them on accounts you own. Every agent ships with permissions scoped to a fixed list of allowed actions, a stop before anything irreversible, and a log of every decision you can read without an engineer sitting next to you.

AI agent. Software given a goal and a set of tools, which decides for itself which tool to use and when, including in situations nobody scripted for it. A workflow follows the steps you wrote. An agent chooses the steps. That one difference is why permissions and audit logs matter more than which model you pick.

What does AI agent development actually involve?

Prompting takes an afternoon. The rest is engineering, in four parts, and three of them are unglamorous.

Scoping the job. We write the task down as a decision: what comes in, what a good outcome looks like, what the agent may conclude on its own, and what it must hand back to a person. If the job cannot be written down that way, it is not ready for an agent, and we say so on the call.

Giving it tools. We connect it to the systems you already run, with its own credentials, at the narrowest access that still does the job: read-only where reading is enough, write access only on the records it has to touch.

Boxing it in. Sandboxed actions, rate limits, an allow-list of operations, and a written security review before it goes live.

Running it. An agent nobody watches drifts, because your process changes and the agent does not know that. We monitor what we ship, read the logs, and adjust. One senior engineer supervises the fleet.

How do you stop an agent taking an action you did not authorise?

Four boring mechanisms, applied every time. They are the difference between an agent you can put in front of customers and one that stays in a demo.

Its own service account. Each agent gets a key of its own. You can revoke it in one click without breaking anything else, and every line in the log names which agent did what.

Permissions scoped in writing. The list of actions an agent may perform is a document you approve. Anything outside the list fails closed and raises a flag instead of improvising.

A gate on anything irreversible. Money leaving, messages sent to your clients, records deleted: those queue for human approval. You can lift the gate later, once you have watched the agent make the same call correctly enough times to believe it.

An audit log you can read. Every input, decision, tool call, result and timestamp, stored in your own database, in language a non-engineer can follow. An agent you cannot audit is not an asset, it is a liability with a good demo.

Anything reading text from outside your company is treated as hostile by default. Text from strangers will try to instruct your agent sooner or later. That work belongs to the security floor, and it is part of the build rather than an upsell.

What can an agent do that a fixed workflow cannot?

Judgment on messy input. A workflow needs a clean form. An agent handles what people actually send.

The shapes this usually takes: reading a supplier email that is half complaint and half new order, and deciding which it is before routing it. Triaging an inbox in English, French, Arabic, Darija and Spanish, and answering the ones with unambiguous answers. Reconciling two systems that disagree, and flagging the rows it cannot resolve instead of guessing. Drafting the follow-up someone would otherwise write at 2am, and leaving it for approval.

The pattern is the same in all of them. The agent handles the part where the next step depends on what the message actually says. Everything either side of that is a workflow, and a workflow is cheaper.

When the judgment depends on what is written in your own files, the retrieval side of the build is RAG pipeline development, which sits on this floor and is usually shipped alongside the agent.

Agents or workflow automation: which does the job?

	Workflow automation	AI agent
Input it expects	Structured: forms, rows, webhooks	Unstructured: emails, chats, documents, voice notes
Behaviour	The same steps every time	Chooses the steps to reach a goal
Cost per run	Close to nothing	Model tokens on every decision
How it fails	Breaks loudly, easy to fix	Acts confidently and wrongly, which is why you need the log
Buy it when	You can write "when X, do Y, then Z"	A step reads "work out what they meant, then decide"

Most systems we ship are both. A workflow catches the message and does the deterministic part, hands the single judgment step to an agent, then takes control back. Deterministic where you can be, intelligent only where you have to be. The longer argument is in AI agents vs. workflow automation. If your problem turns out to be rules all the way down, we build it as workflow automation and charge you less.

Who this is for, and who it is not for

Good fit	Poor fit
Your team spends its day reading unstructured input and deciding what to do with it	The work is genuinely rule-based, in which case buy a workflow and keep the difference
Someone is doing judgment work at an hour nobody should be working	You want people removed entirely; every agent we build has a supervisor by design
You already run n8n, Make or Zapier and keep hitting the wall where the rules run out	You need something impressive for a board meeting next week
Your front line answers in more than one language	You cannot say what a wrong action would cost, and that number decides where the gates go
You have to explain to an auditor or a client why a decision was made	You would rather someone else held the credentials

What you own when the agent ships

Your cloud, your repos, your keys. The agent runs on infrastructure with your name on the account, and you can log in and read the code today, without asking us.

Anything that can be open source and self-hosted is, so no licence line appears on any invoice. Model APIs are billed to your account at cost, with no markup, so scaling the agent does not quietly turn into per-seat rent. The reasoning behind that default is set out in own your automation stack, and it holds for everything else we run.

Where your data sits, and where we sit

Our team sits in Casablanca. Your data does not have to be. Agents deploy to whichever region you require, EU-region hosting included, with a data processing agreement and standard contractual clauses where transfers apply, and zero-retention settings on model providers that offer them. What we hold as a company is written out in our privacy policy, and the contractual side sits in our terms.

Morocco is the 55th state party to the Council of Europe's Convention 108 on data protection, in force since 1 September 2019, which is usually the first thing a European legal team asks about. It is not an EU adequacy decision, and we will not present the convention as a substitute for one. Bring your obligations to the call and we design the deployment around them rather than around our own convenience.

How the first two weeks go

Thirty minutes on a call: you describe the workflow that hurts, we map which parts are rules and which parts need judgment. A priced roadmap lands in your inbox in writing within 48 hours, and it is yours to keep whether you sign or not.

If you do sign, the first deliverable is something working in your hands within two weeks. Those two weeks are refundable, no conversation needed. The work is founder-led by Sami Khalfi and Amine Lamfichekh, in English, French, Arabic or Spanish, wherever your team is. What working from a Casablanca base changes in practice is set out on our AI automation agency in Casablanca page.

Bring the process that is eating your week. If an agent is the wrong answer, we tell you what the right one is. If you want the wider picture first, walk all five floors, or read the longer write-ups on the blog.

Common questions

What is AI agent development?

AI agent development is building software that is given a goal and a set of tools, and decides for itself which tool to use and when. Unlike a workflow, which repeats steps you wrote, an agent handles situations nobody scripted. That autonomy is why a serious build spends most of its effort on scoped permissions, sandboxed actions and audit logs.

How long does it take to get an AI agent running?

Your first working deliverable lands within two weeks, and those two weeks are refundable with no conversation needed. Before that, a thirty-minute call maps which parts of your process are rules and which need judgment, and a priced roadmap reaches your inbox in writing within 48 hours. The roadmap is yours to keep whether you sign or not.

How do you stop an AI agent taking a wrong action?

Four mechanisms, on every build. The agent gets its own service account, so you can revoke it in one click. Its permitted actions are a written list, and anything outside it fails closed. Irreversible steps such as payments, client messages and deletions queue for human approval. Every decision is logged in your own database, readable by non-engineers.

How much does AI agent development cost?

There is no price list, because no two companies hurt in the same place. Scope is mapped on a thirty-minute call and priced in writing within 48 hours. Running costs stay transparent: anything open source is self-hosted with no licence line, and model APIs are billed to your own account at cost, with no markup.

Who owns the agent once it is built?

You do, from day one. The agent runs on your cloud, in your repos, under API keys with your name on them, and you can log in and read the code without asking us. The handover is documented so the next engineer can take over without a phone call. Nothing important lives on an account only we can reach.