AI Agent Development
We design and build business copilots, tool-calling agents and customer support agents that look things up, take approved actions in your systems, and hand over to a person when they should.
Prefer chat? Message us on WhatsApp
For business leaders weighing where agents fit, IT teams who will have to secure them, and operations managers whose teams will work alongside them.
Chatbot, copilot or agent?
The word "agent" covers very different systems. The more an assistant is allowed to do on its own, the more engineering, testing and oversight it needs. We help you pick the least autonomous option that solves the problem.
Answers
Responds to questions from approved content. Takes no actions.
Assists
Works beside a person: drafts, summarises, looks up records. The person decides and clicks.
Acts
Calls your APIs to complete steps on its own, within permissions, with approvals for sensitive actions.
If every step of the job is known in advance, a rules-based workflow is usually cheaper and more predictable than an agent. That is the territory of AI Automation.
AI agent architecture
A demo agent is a prompt and a model. A production agent needs every part below. Select one to see what it does and why your IT team will ask about it.
Part 1 of 8
The role, goals, tone and hard rules the agent works to. Written like a job description and versioned like code.
Part 2 of 8
The reasoning engine. We pick the model per task, balancing quality, speed, cost and where your data may be processed.
Part 3 of 8
Approved documents and records the agent can search, so answers come from your sources rather than the model's memory.
Part 4 of 8
Named functions the agent may call, such as "look up order" or "create ticket", each mapped to one of your APIs.
Part 5 of 8
Scopes that decide which tools and records the agent can touch, often narrower than the person it is helping.
Part 6 of 8
Checkpoints where a person confirms an action before it happens, such as a refund, a price change or an outbound email.
Part 7 of 8
What the agent keeps between steps and conversations. We keep it minimal and expire it on a schedule you approve.
Part 8 of 8
A record of every tool call and answer, plus test sets that are rerun whenever instructions, tools or models change.
Agents we build
Each type below includes the signals that it fits your situation and the signals that something simpler would serve you better.
An assistant inside the tools your team already uses that prepares work for a person to finish: account briefs before a sales call, a first draft of a proposal, a summary of a long ticket thread.
Good fit when
People repeat the same research or drafting every day, and a human must stay accountable for the result.
Think twice when
The output goes straight to customers with no review, or the inputs live in systems with no API.
Agents that complete multi-step tasks by calling your APIs: check stock, update a CRM record, book a slot, raise a ticket. Every tool has a defined scope and every call is logged.
Good fit when
The task needs judgement about which step comes next, and the systems involved expose reliable APIs.
Think twice when
The steps never change. A fixed workflow will be cheaper and easier to audit.
Agents that answer employee or customer questions from approved documents and records, cite their sources, and say "I don't know" when the answer is not there.
Good fit when
Answers exist in your documents but are slow to find, and the content is kept reasonably up to date.
Think twice when
The source documents conflict or are years out of date. Fix the content first.
Web chat, WhatsApp or voice agents that resolve routine requests such as order status, returns policy or appointment changes, and route everything else with a summary attached.
Good fit when
A large share of contacts are repetitive, and you have clear policies on what can be resolved without a person.
Think twice when
Most contacts are complaints or complex disputes that need empathy and discretion.
Several narrow agents coordinated by an orchestrator, for example one to research, one to draft and one to check against policy, with a person approving the final step.
Good fit when
A single agent becomes unreliable because the task mixes very different skills or data sources.
Think twice when
A single well-scoped agent already performs well. More agents means more cost and more to test.
Tool calling, explained
Tool calling lets a language model ask your software to do something specific. The model never gets direct database access. It can only request the named tools we expose, and our code checks each request against the agent's permissions before anything runs.
Where is my order 48213? It was due on Tuesday.
orders.lookup({ id: "48213" })
→ status: shipped · carrier: delayed at hub
shipping.track({ order: "48213" })
→ new estimate: Friday
refunds.create({ order: "48213", reason: "damaged" })
→ paused · sent to support lead with photos and order history
Your parcel is delayed at the carrier hub. The new delivery estimate is Friday, and I will message you if that changes.
Tool names, order details and timings are invented to show the pattern. Real tools map to your own APIs.
Controlled autonomy
We agree an autonomy level for each action, not for the agent as a whole. Most agents launch at level 1 or 2 and move up only when logs and evaluation results justify it.
Level 0 · Suggest
The agent recommends. A person does everything.
Oversight effort required: Lowest
Example: customer support
Suggests a reply and the relevant help article. The support agent writes and sends.
Level 1 · Draft for approval
The agent prepares the action. A person approves each one.
Oversight effort required: Low
Example: customer support
Drafts the reply and the CRM update. One click sends both, or the agent edits first.
Level 2 · Act within limits
The agent acts alone on low-risk tasks and asks for anything outside its limits.
Oversight effort required: Medium
Example: customer support
Answers order-status questions alone. Refunds and complaints go to a person.
Level 3 · Act and report
The agent acts on a defined task and people review samples and exceptions.
Oversight effort required: Needs strong evaluation
Example: customer support
Tags and routes all incoming tickets. A lead reviews a daily sample and the exceptions queue.
For your team
Agents fail most often because one of these groups was not part of the design. We plan for all three from the first workshop.
The question they ask first
"Is this worth doing, and what could go wrong?"
What we give them
The question they ask first
"How does it authenticate, what can it reach, and where is data processed?"
What we give them
The question they ask first
"What changes for my team on Monday morning?"
What we give them
Agent readiness check
Tick what is true for the task you have in mind. Nothing is sent anywhere; this is a quick way to see where a discovery call would start.
0 of 6 in place
Start with discovery. We would map the task and data first, and may recommend a copilot or a simple workflow before a full agent.
A good candidate for a scoped pilot at autonomy level 1, with the gaps closed during the build.
Well prepared. We can move quickly to architecture and a prototype tested on your real tasks.
Talk it through with usQuestions
Start with one agent
Describe the task, who does it today and which systems are involved. We will tell you honestly whether an agent, a copilot or a simpler workflow fits best.