AI agent development: software that does the task, not just the talking

Agent or workflow? A five-question test

Most tasks people call agent work are better served by a fixed workflow with one model call inside it. Cheaper, easier to test, easier to explain to an auditor. Use these questions before you commission anything.

If most answers point to workflow, see AI automation. The difference is also explained in AI agent vs chatbot.

What an agent we build is made of

Every agent has the same five parts. Naming them early keeps the project honest.

• Goal and scope: one sentence on what the agent is for, and a list of what it must refuse.
• Tools: the actions it may take, each wrapped as a function with its own permission level.
• Knowledge: the documents or records it may read, kept current by a sync job.
• Memory: what it keeps between steps and between runs, and what it forgets.
• Controls: spending limits, timeouts, approval gates and a complete action log.

A worked example, described in general terms: an agent that handles supplier invoice questions reads the email, finds the purchase order, checks payment status in the accounting system and drafts a reply. It may read everything and send nothing. A person approves each reply for the first weeks, and approval is lifted only for categories where the edit rate stays low.

Permission levels we use

Every agent starts at level 1 or 2. Moving up is a decision your team makes after reading the logs, not something we assume.

Our process

1. Scope the taskWe shadow the task or read 30 to 50 past cases and write down the steps, the exceptions and the failure that would hurt most.
2. Build an evaluation setA spreadsheet of real cases with the right outcome for each. This is how we know whether version two is better than version one.
3. PrototypeA working agent in a sandbox with fake or copied data. You can try it within the first weeks of the project.
4. HardenAdd permissions, rate limits, fallbacks when a tool is down, prompt injection defenses and the action log. We use our AI agent security checklist.
5. PilotRun alongside your team on live work with approval on every action. Measure edit rate, time saved and errors.
6. Hand overCode, prompts, test set and runbook go to you. Optional monitoring keeps running after launch.

What you get and how long it takes

• A working agent connected to your systems through scoped credentials.
• The evaluation set and a script to re-run it after any change.
• An action log you can read without a developer.
• A short runbook: what to do when the agent is wrong, a tool is down or a model is retired.
• Written notes on data handling for your privacy review, covering GDPR, UK GDPR or CCPA as they apply to you.

A single-task agent usually reaches a supervised pilot within a few weeks. Agents that touch several systems or need data cleaning take longer. Your quote names the stages and the dates for your scope. Model, hosting and tool costs are listed separately so you can see what the agent costs to run.

How we report and where our experience sits

During the pilot you receive a weekly note with four numbers: tasks handled, edit rate, errors caught by approval and cost per task. Those come from the agent's own log, which you can open.

We have not published an agent case study yet, so we will not pretend one. Our closest real work is Aura Realty, where we built a custom CRM that pulls leads from the website, Instagram, WhatsApp and calls into one place. That is the sort of system an agent would sit on top of.