Agentic AI
Agents will not replace your specialist software. They will operate it.
- An agent can now drive specialist software through its screens, as a trained person would. A solo founder used Codex to operate a professional 3D fashion tool she had never learned.
- For companies built on specialist or legacy tools, we think the near-term win is an agent that operates the tools they already have, not a rebuild.
- The software keeps its rules, formats and validation. The agent does the clicking and typing, and a person checks what comes out.
- Start with one task. Run the agent and a person on the same cases, keep the better result, and let the numbers decide what to automate next.
We think a company that runs on specialist or legacy software should put an agent in front of the tools it already has, with a person checking the output, before anyone proposes a rebuild. Those tools hold the rules, formats and checks a rebuild would have to recreate, and the hours spent clicking and re-keying around them are the part an agent can now take over.
A founder ran 3D design software she never learned
On 17 August, Lenny's Newsletter published an episode of its How I AI podcast in which host Claire Vo talks with Yana Welinder, a former product leader who is building a fashion brand alone. Welinder uses OpenAI's Codex with computer use, meaning the agent reads the screen and clicks and types the way a person would, to operate design software she has never learned.
The episode's weekly roundup says Codex drove CLO, a professional 3D fashion design tool, to produce the CAD, or computer-aided design, files she needed for 3D printing.
The useful detail is the split. Codex could not make a finished CAD file by itself, and Welinder had never learned CLO, yet together they produced the files. The roundup concludes that software sold as a service is not going away and is "gaining a new kind of user."
We read this as a pattern for any operation that depends on specialist software, such as claims systems, order and inventory screens, lab software or engineering tools. The agent brings general reasoning. The tool brings the domain rules, the validation and the file formats it already encodes.
In why general models keep beating specialized AI we argued that a general model with the right tools beats a model trained for one niche. The specialist software a company already owns is one of those tools.
Why we would not start with a rebuild
A rebuild means recreating every rule, edge case and file format the old tool handles, moving its data and retraining the people who use it. In our view the costly work usually sits around the tool, in the copying, re-keying and checking between screens. An agent that operates the tool takes on that work and leaves the system of record where it is.
Older specialist software often has no API, the interface other programs use to drive it. Microsoft aimed its computer use launch for Copilot Studio at that gap in April 2025, with agents that click buttons, select menus and type into fields when no API is available. Its summary was "If a person can use the app, the agent can too." Microsoft also says the agent adapts when screens change, which is a vendor claim to test on your own software.
We would still use an API or a connector wherever one exists, and keep screen control for software that offers neither. In our three-layer map of agent tools, both belong to the crew layer, the part that lets a model act. The rule from that post holds here. Pay for safe action, not for the ability to click.
Code has a test suite. Most operations software does not.
Software teams already let agents work with less supervision. On 19 August, Cursor said its cloud agents can now pick up work on their own, from a pull request, a Slack thread or a schedule, and keep working toward a goal until it is met. A pull request is a proposed code change. Cursor wants agents that build and ship software "without the need for intervention at each loop."
That can work in code because every change runs through CI, the automatic tests and checks that run on each change. Cursor's own example goal is to fix all flaky tests, the ones that pass and fail at random, and make CI green. A goal like that rewards whatever turns the checks green, and METR has documented frontier models trying, often successfully, to raise their scores by modifying the tests or the scoring code.
A claims screen or a CAD tool usually has no test suite behind it, so the first check on the agent's work is a person. Each correction that person makes is the raw material for an automatic check later.
The test automation vendor Tricentis is building this check layer for agents. On 20 August, VKTR reported that Tricentis had announced AgentScore, designed to grade an agent's behavior in real workflows and recommend whether to ship or block it. It also announced Aida, an agent that explores web and desktop applications to find defects. Both come from Tricentis Labs, which gives enterprise users early access. For operations work we would use three outcomes, with a person in the middle.
- Work arrives: an order, a claim, a drawing. Leads to Agent runs the tool.
- Agent runs the tool: screens, connector or API. Leads to Automatic checks.
- Automatic checks: record saves, totals match. Leads to Accepted (pass), Person reviews (unsure), Blocked (fail).
- Accepted: every check passes.
- Blocked: a check fails, done by hand.
- Person reviews: checks cannot settle it. Leads to Agent runs the tool (corrections).
Run the agent beside a person first
One of Welinder's hardest remaining problems is turning designs into accurate sewing patterns. According to the roundup, she has human patternmakers and Codex work on it at the same time, compares the results and keeps the better one.
We think that side-by-side comparison is the cheapest honest test an operator can run on specialist software. It costs some duplicated effort, and it gives a direct comparison on real cases before anyone has to trust the agent.
How to start on one tool
Pick one task inside one tool
Choose a task a person does many times a week, with an output someone can check, such as a saved record, a file that opens or a total that matches.
Write the task down as a spec
List the inputs, the steps, the exceptions and what done looks like, as you would for a new hire.
Give the agent its own login
Grant the least access the task needs and log every screen and action. Copilot Studio, for one, keeps screenshots and reasoning steps for each run.
Run the agent and the person on the same cases
Compare the outputs case by case and keep the better one.
Write the verdict rule
Decide which outputs are accepted after automatic checks, which go to a person and which are blocked.
Turn every correction into a check
Each case the person fixes becomes a test the agent must pass before any change to its instructions, its model or the tool's version.
Where we land
The rebuild question does not go away, but data can answer it. If the agent keeps failing on the screens, or the person rejects most of its work, that is evidence for an API or a replacement. Until that evidence exists, keep the software and change who operates it.
Our AI process automation method follows the same order: map the workflow, automate it across the tools you already have, and route the exceptions to a review queue your team owns.
Pick one task in one tool this week, write down the checks that would prove it was done right, and run the agent beside the person who does it today.
Sources
- Lenny's Newsletter, How a solo founder used Codex and ChatGPT to launch a fashion brand without engineers and its weekly roundup, 17 August 2026
- Cursor, Cloud agents and Cursor harness improvements, changelog, 19 August 2026
- VKTR, Tricentis debuts 3 AI agents for quality engineering, 20 August 2026
- Microsoft, Announcing new computer use in Microsoft Copilot Studio for UI automation, April 2025
- METR, Recent frontier models are reward hacking, June 2025