Automation
Intelligent automation software that reasons, writes code and asks first
Intelligent automation is usually sold as rules with a model bolted onto one step. AgentCode puts the reasoning where the work is: the agent reads your code and your description, writes the automation, and waits for a person to approve it.
You describe the work · the agent writes the code · your team reviews before anything goes live
In short
Intelligent automation software is the label for automation that does more than follow a fixed script. It reads documents, interprets messages, decides between paths, and hands the unclear cases to a person. In most products the intelligence is a component, a model called from one step in a flow. In AgentCode the intelligence is the builder itself. An agent reads your existing systems, works out what you asked for, writes the code, runs your tests, and gives the change to a person.
What intelligent automation looks like here
- Documents read and turned into structured records your systems accept
- Free text messages sorted into the categories your team already uses
- Rules written as code, so any decision can be explained afterwards
- A plan you read before the agent touches a single file
- Your own tests run against every change the agent proposes
- A human approval gate on the path to production, every time
A process from IT
The problem
Service desk tickets arrive as free text, so somebody reads every one and decides which team it belongs to before any actual work on the ticket starts.
What the agent builds
The agent builds a triage app that reads the ticket, proposes a category and an owning team, and puts the ones it is unsure about in front of a person.
What you end up with
Tickets that arrive already sorted, a short review screen for the unclear ones, and the classification rules as readable code in your repository.
What intelligent automation means in practice
The term covers the point where automation stops being a script. A rule based automation needs every case to look the same. Intelligent automation handles the cases that do not: an invoice in a layout nobody has seen, an email that contains a request and a complaint at once, a form filled in by a person in a hurry.
- Document understanding: pulling fields out of files that vary in shape
- Classification: deciding what a message is about, so it reaches the right queue
- Judgement with a threshold: acting when confident, asking a person when not
- Orchestration: joining the interpreted steps to the deterministic ones around them
The last point is where most projects stall. A model that reads invoices is useful only once something writes the result into the ledger, shows the doubtful ones to a person, and keeps working next month. That surrounding software is ordinary code, and it is exactly what the agent writes for you.
The agent is the builder, and the judgement inside what it builds
Two kinds of intelligence appear in this product, and it helps to keep them apart.
The first is the agent that builds. You describe what your department needs, and it reads your repository, writes a plan you can read, changes the files, runs your tests and stops. A person approves before anything merges, every time, with no exception for small changes.
The second is the judgement inside what gets built. When a process genuinely needs interpretation, such as reading a supplier document or sorting inbound messages, the agent writes that step into the application together with its fallback: a confidence threshold, a review screen, and a record of what was decided and why.
- You can open the file and read the code behind every automatic decision
- Uncertain cases reach a person by design, not by accident
- None of your business logic is hidden inside a product you cannot inspect
Read how it works for the full sequence from description to approved change.
Where IT teams put it first
IT teams meet intelligent automation twice: once as the department asked to support automation built by another team, and once as a team with its own queue of repetitive work. Both are sensible starting points, and both benefit from rules that are written down rather than remembered.
- Ticket triage: categorising and routing inbound requests before a person reads them
- Access requests: checking a request against policy and assembling the approval trail
- Log and alert summaries: turning a noisy feed into the few items worth acting on
- Onboarding checklists: accounts, hardware and permissions tracked as a single case
- Shadow spreadsheets: replacing the critical file somebody maintains by hand with a small app
That last one is common enough to have a page of its own: see internal tools. For the purely deterministic side of the same work, read RPA software, and see pricing for plan limits.
Questions teams ask
Is intelligent automation just RPA with a model attached?
In many products that is a fair description: a flow with one step that calls a model. Here the split is different. The reasoning happens while the automation is being built, so what runs afterwards is code you can read, with interpretation used only where the process truly needs it.
How do we know why an automatic decision was made?
Because the decision path is written down as code in your repository, and the cases the agent was unsure about go to a review screen rather than through unchecked. You can open the file, read the rule, and change it by describing what should happen instead.
Can the agent change things without us noticing?
No. The agent writes a plan, edits files and runs your tests, then stops and waits. Merging is a human action, always. Nothing reaches your live systems until somebody with the right access has read the change and accepted it.
Do you train on our code or our data?
No. Your code stays in your repository and your data stays in your systems. We do not train on either. The agent reads what it needs in order to complete the task you described, and the output is a change you review like any other change to a business system.
More the agent builds
Give the agent your messiest queue
Describe how the cases should be sorted, then review the app that sorts them.