Agentic systems and AI agent development

Build an agentic system for your existing software.

An agentic system handles work where the input changes and the next action cannot be fixed ahead of time.

Its AI agent chooses an allowed action, sees what happened, and chooses again until the job finishes or has to stop.

I build the applications and integrations around it. I add repeatable evaluations, monitoring, and approval where the workflow needs it.

Known rules go in code. I use a model for interpretation and leave important decisions to a person.

Tell me about the workflow Do you need an agent?

Agent or ordinary automation?

I use ordinary software when every step is known.

An agent is for work where the result is clear but the next action depends on what happens during the run.

Agent: The model chooses the next allowed step or tool and repeats until the job is done or reaches a stop condition.
Agentic workflow: Some stages stay fixed in code. The model chooses what happens between them.
LLM-powered step: The model reads or writes one thing. Code decides what happens next.
Fixed workflow: Code follows known rules from start to finish. It does not need an agent.

Before I choose a model

First, I write down the job.

I look at who does the job, what comes in, and what counts as done. Then I decide which actions software can take and which need a person.

What the agent may do

I define what counts as success, which tools it can call, when it must stop, and what requires approval.

Access to existing systems

The agent gets its own identity and only the permissions it needs. The application saves state and checks what comes back from files, APIs, and other software.

When a run cannot finish

I cap retries and run time. If the model cannot finish, the application saves its place and hands the work to a person.

I build this work in .NET/Azure or TypeScript/Node.js. I pick the framework after I understand the job and the software it must connect to.

Testing and monitoring

I test the whole run.

I check the chosen action, tool results, final state, and why the run stopped. After I fix a failure, I rerun it as a test.

Repeatable evaluations

I rerun a fixed set of representative examples and compare each result with the expected one.

See the run

I want to see what came in, which actions it chose, and what each tool returned. The run record also shows approvals, errors, timing, and why it stopped.

Human approval

Before an important action, the system pauses. The person responsible sees the evidence and decides whether it continues.

Failure and recovery

If a tool fails, the agent may retry within a limit. If its state is uncertain, it stops before taking another action.

Client work

Two client systems that included an AI agent.

Controlled order workflow

The agent read emails and PDFs and kept choosing allowed actions until done or stopped. Code checked results. I reran tests, reviewed each run, and operators approved exceptions.

See the order workflow

Supervised shipping agent

Fixed automation handled known ERP screens. On unexpected ones, the agent used AI vision to choose actions. I reran the tests. A supervisor watched and could step in.

See the shipping application

View all projects

Release and handoff

The next engineer should not need me to run it.

Before release, I limit what the agent can do and save enough detail to review each run.

I hand over the code, integrations, tests, and run history.

The notes explain how to pause the agent, recover a run, and change the system.

Fit

I’ll tell you when fixed automation is enough.

Bring me one workflow with changing inputs, a clear result, and someone who can approve the risky actions.

What keeps changing in the workflow?

Send me one real input and tell me what the system should do with it. Mark anything that still needs a person.

Tell me about the workflow