AI
AI agents
An agent is just a model that has been allowed to use tools and run several steps in a row. That sounds more advanced than it is, which is both good and bad news.
- Same-day reply
- Fixed price
- No lock-in
- Bounded tasks
- Logged at every step
- Human in the loop
- Interruptible
Service
AI agents
What they are good at
Agents work best on tasks where the steps are known but the order depends on what turns up along the way. Investigating a support case across three systems. Assembling a draft from several sources. Cleaning up data typed in by humans over ten years.
They work badly on tasks where there is no right answer, or where an error will not surface until much later. There, a model without judgement is a poor employee.
How I bound them
- One agent, one job. An agent that can do everything does nothing properly.
- Read tools are cheap to hand out. Write tools require a decision about what may happen without approval.
- Everything is logged, both what the agent did and why it thought it should.
- There is always a way out: a step limit, and a way to stop the whole thing.
On keeping a human in the loop
The most useful setup I have built is rarely the most autonomous one. It is the one where the agent finishes the work and puts it up for approval, so a human spends two minutes instead of forty.
That model is also easier to start with, because nobody has to dare let it loose on day one. Trust can be built once you can see what it does.
Deliverables
What you get
Concrete deliverables, not a report of recommendations someone else has to find time for.
- One bounded agent
With a clearly defined remit, a step limit and a way to stop it midway through a task.
- A tool layer
Access to your systems, split between reading and writing. Write access only arrives once the read side has run for a while.
- A full log
Every step, every tool call and the reasoning behind it. Without it you cannot debug, and you cannot answer what happened.
- An approval flow
Where it matters, the agent queues its work for approval instead of carrying it out itself.
Process
How it works
Four steps. You know what happens when, and you can stop after any of them.
- 01
Find the task
We look for the work that gets done many times a week, follows a pattern and annoys the people doing it. That is where the first win is.
- 02
Measure the baseline
How long does it take today and how often does it happen. Without that number you cannot say afterwards whether anything improved, and it becomes one gut feeling against another.
- 03
Build and evaluate
I build a test set of real examples early and measure against it every time the prompt or model changes. It is the single thing most often missing in projects I take over.
- 04
Put it into production
Into the systems people already work in, with logging on everything and a way to switch it off. Plus an agreement on who owns it once I am gone.
Who does the work
I am the person you will be working with
My name is Mikkel and I have been building for the web since 2010. There is no project manager in between, and you will not get a junior on the job because the senior was busy. What you agree is what gets built, and I am the one building it.
Day to day I am a partner at Bigum&Co, and I have worked on more than 400 projects across ecommerce, pharma, construction and the public sector. Call or write — I usually reply the same day, and I will say straight away if the job is not a fit.
- Experience
- 16 yrs
- Projects
- 400+
- Rating
- 5.0
- Reply time
- Same day
In depth
AI agents
What an AI agent is, and what it is not
The word covers something fairly simple: a model given access to some tools and permission to use them several times in a row until it thinks the task is done. That is the whole difference from an ordinary chat. There is no hidden intelligence underneath.
That makes them good at tasks where you cannot write the order down in advance. To work out why a particular order was not delivered, you do not know beforehand whether the answer sits in inventory, with the carrier or in an email. A person checks those three places in whatever order makes sense given what turns up. An agent can do the same.
It makes them bad at tasks where there is no right answer, or where an error only surfaces months later. Prioritising between two customers, making a hiring decision, judging whether a piece of copy hits the tone — there a model without judgement is a poor employee. It will answer, and it will sound certain.
The trap people fall into most often is building one agent that does everything. It becomes impossible to test, because the number of possible paths explodes, and impossible to debug, because you cannot tell which step went wrong. Several small agents with one responsibility each are duller to talk about and markedly easier to live with.
How trust in an agent gets built
Nobody lets an agent loose on their order system on day one, and they should not. The order that works is the same every time: first it may only read, and the result is shown to a person. Then it may propose an action a person approves with one click. Finally it may carry out the actions where mistakes are cheap to undo.
Along the way the log builds into something you can make decisions from. After three weeks you can see how often the agent was right, where it was unsure, and which case types it consistently gets wrong. Those case types simply get routed to a person. That is a far better answer than trying to make the agent good at everything.
There are tasks where the approval step should never disappear. Anything that sends something out of the building to a customer, and anything that moves money, belongs there. That is not a technical limitation, it is a decision about who answers for the outcome.
Reviews
What people say about me
Mikkel is a man after my own heart. Executes at lightning speed, has an eye for detail and is always there when you need him. It does not get better than that.
Tini Owild(translated from Danish) He is simply a pro. Highly competent and precise, brings good input, and is a huge support.
Carsten Johan Thessen(translated from Danish) I have never come across a developer as dedicated and capable as Mikkel.
Jonas Krogslund(translated from Danish)
Services
Why me
- 16 years of experience
I have built for the web since 2010 and seen most of the ways a project can come off the rails. It makes the estimates more honest.
- 400+ projects
Across ecommerce, pharma, construction and the public sector. Chances are I have seen something close to your job before.
- You talk to the person coding
No project manager in between, and no junior on the job because the senior was busy. What you agree is what gets built.
Brands I have worked with
A selection. Some freelance, some through agencies, some across several years.
Questions
Frequently asked questions
- What if the agent does something wrong?
- Then it has to be detectable and reversible. That is why I start with read access and add write access afterwards, one action at a time, with logging on all of it.
- Can it work inside our own systems?
- Yes, if the systems have an API. If they do not, a layer can often be built in between, but it belongs in the estimate because it is frequently the largest part of the job.
- How long does it take to build?
- A bounded agent on one system is typically two to four weeks including testing. Working across several systems, it is the integrations that set the timeline, not the AI part.
- Which systems can an agent work in?
- Anything with an API. Shopify, Google Workspace, HubSpot, e-conomic, Jira, your own system. If there is no API, a layer can often be built in between, but it belongs in the estimate because it is frequently the largest part of the job.
- What does it cost to run an agent?
- The price per task depends on how many steps the agent takes and how much text goes along each time. I work out an estimate early, because it often changes how the solution should be built.
- Can the agent work outside office hours?
- Yes, and that is often the point. Cases that arrive in the evening can be investigated and prepared by the time someone starts in the morning. The decision itself can still sit with a person.