AI Agents
An agent decides its own steps instead of following a fixed script. That makes it useful for work that varies, and it makes the permissions around it the most important part of the build.
What an agent actually is
Strip away the terminology and an agent is a model that has been given a goal, a set of tools it can call, and permission to decide the order in which to use them. Instead of running step one, then step two, then step three, it looks at the situation, chooses an action, sees the result, and decides what to do next.
That loop is the whole difference. A conventional automation cannot handle an input it was not designed for. An agent can attempt it, which is genuinely useful and also the reason it needs boundaries that a scripted workflow never does.
Where agents beat a fixed workflow
The deciding question is whether the steps vary between cases.
Writing a product description is a good example. For a simple accessory, the spec sheet is enough. For a technical product, the useful description needs the comparable models, the compatibility notes and possibly the support documentation, and which of those matter differs per item. Encoding that as a decision tree is possible but miserable to maintain. An agent works out what it needs for the item in front of it.
The same applies to analysis where the interesting question depends on what the data shows, and to operational chasing where the right next action depends on what has already happened. Where the path is genuinely fixed, an agent is the wrong tool and we will recommend a plain automation instead.
Permissions are the design
The most important decision in an agent project is not which model to use. It is what the agent is allowed to touch.
We define this as an explicit list: these data sources it can read, these actions it can take, these it must never take without a person approving. An e-commerce agent might update a draft description freely, propose a category change for review, and have no ability at all to change a price. A support agent might read the whole ticket history and never issue a refund.
The uncomfortable version of this question is what happens if the agent is confidently wrong, because eventually it will be. Designing around that is not pessimism, it is the difference between a system you can leave running and one that needs somebody watching it.
Testing something that varies
An agent will not produce identical output twice, which makes conventional testing a poor fit. What works is an evaluation set: a collection of real cases with known good outcomes, run repeatedly, scored on how often the agent gets there.
That gives you a number to compare against when you change a prompt, swap a model or add a tool, which matters more than it sounds. Without it, tuning an agent is guesswork, and it is very easy to make a change that improves the case you were looking at while quietly breaking three you were not.
We also keep traces of every run in production. When output is wrong, the trace shows which step went wrong, and that is usually a fixable problem in a tool description or a missing piece of context rather than something fundamental about the model.
Start supervised, then loosen
The sensible rollout is narrow and watched. One process, a limited tool set, and a person approving actions while evidence accumulates about where the agent is reliable and where it is not.
That evidence is what autonomy should be based on. After a few weeks you will typically find some action types the agent handles better than expected, and one or two it consistently gets wrong in a specific way. You relax approval on the first group and either fix or permanently gate the second. It is slower than switching everything on, and it is the reason the system is still running six months later.
What you get
Content agent
Produces articles, product descriptions, video outlines and newsletters, choosing its own approach per piece and checking its output against your brief before handing it over.
Analytical agent
Pulls data from your tools, works through it and reports what changed and why it might matter. Useful where the questions repeat but the answers do not.
E-commerce agent
Updates descriptions, tidies product data and proposes changes to categories or attributes, flagging what it is unsure about rather than editing silently.
Operations agent
Plans tasks, tracks statuses, chases overdue items and keeps project data current across the tools your team already uses.
Multi-agent workflows
Several agents with narrower jobs handing work between each other and into your CRM, ERP or CMS. Narrow roles are easier to evaluate than one agent that does everything.
Guardrails, traces and approval gates
Scoped permissions, spend limits, a full record of every step the agent took, and defined points where a person signs off before anything irreversible happens.
How we work
- 01
Pick a process with genuine variation
Agents earn their complexity where the steps differ per case. If the path is the same every time, a plain automation is cheaper, faster and easier to trust.
- 02
Define the tools and the permissions
We list exactly what the agent may read and what it may change. This list is the security boundary, and it is deliberately narrow at the start.
- 03
Build an evaluation set
Real cases with known good outcomes, including the awkward ones. Without this you have no way to tell whether a prompt change improved anything or just moved the failures.
- 04
Run supervised first
Every action goes to a person for approval while you build up evidence about where it is reliable. Autonomy is granted per action type, based on what the traces show.
- 05
Monitor cost, quality and drift
Token spend, step counts, failure patterns and output quality tracked after launch, because model updates and changing inputs both move behaviour over time.
Tools and technology
Where a solid open-source tool exists, we choose it over a closed one. No lock-in to a single vendor, and costs you can actually predict.
- Vercel AI SDK
- LangChain
- LangGraph
- Model Context Protocol
- n8n
- Ollama
- Next.js
- PostgreSQL
- Supabase
- OpenAI
- Anthropic Claude
- Google Gemini
Frequently asked questions
More services in this category
Basic AI Implementations
Not every AI project needs to be a platform. The fastest return usually comes from a few narrow tools that take one repetitive task out of someone's week.
No-Code and Low-Code AI
We design and build the logic; you keep the ability to change it. Workflows you can open, read and adjust yourself, without waiting in a development queue.
Process Automation
If someone does the same thing the same way for the third time, it probably should not be manual. We automate the repeatable part and leave the judgement where it belongs.
Read testimonials from companies that trusted us
Delivered well ahead of the deadline.
ZanReal's individual approach is impressive.
Knowledge and business intuition make them a valuable partner.
Quick solutions that reduced costs by 99%.
Latest
Technical guides, security insights, and what we've learned building with AI.
Have a process where the steps change every time?
Message usDescribe how it runs and where the cases differ, and we will tell you whether an agent is justified or a plain automation would serve you better.
Can't keep up with changes in AI world?
Let us do the heavy lifting. Every week we distill the most important AI developments into a focused 5-minute briefing — so you stay ahead without the noise.
Find out more