Support team builds its own report with AI agents in 17 hours on a $20 subscription
A support lead kept asking for new targets by group, then a different period, and every report change landed in the development queue. The team described the task in plain words in SimpleGen, an AI tool on the SimpleOne platform, and the agents delivered a finished page in 17 hours. By the product owner's estimate, doing it manually would have taken at least twice as long. All 786 330 Claude Code tokens fit within a $20-a-month subscription. A human still checks the formulas.
AI-processed from Habr AI; edited by Hamidun News
A familiar situation: a support lead gets a report on tickets and is happy with it. A week later they ask for targets for each group, then a different period, then one task type broken out separately. Each request takes five minutes, but development has a queue for it, and in SimpleOne's built-in report builder every variant would have to be assembled separately.
The team decided not to wait in line. They described the task in plain words in SimpleGen, an AI development tool on the SimpleOne platform, where companies run their support processes. Here, "harness" means the working scaffolding in which several AI agents write, check and ship code.
Getting from request to finished report took 17 sessions and about 17 hours, including pauses. By the product owner's estimate, doing the report manually would have taken at least twice as long. The entire token spend, 786 330 tokens in Claude Code, fit within a $20-a-month subscription.
For you, this answers the question of whether small internal requests can be handled without a separate development project.
Why the built-in report builder wasn't enough
SimpleOne's standard reports and dashboards are good at summing up tables. But flow balance (how many tickets came in and how many were closed), median response time, the share of reopened tickets and targets by group all need separate calculation logic. On top of that, the history table, by the team's estimate, grows to roughly 30 million records. As a result, line managers, heads of service departments and IT directors page through ticket lists by hand and miss critical items. BI exists for annual statistics, but support needs its own metrics, targets and screen layout.
How the four stages work
SimpleGen uses different agents: a code author, a fixer, a reviewer and others. Time was distributed unevenly across the stages:
- spec: about 5.5 hours, with 11 "write and review" iterations and 28 questions from the agent;
- plan: 1.5 hours, with 11 phases and 42 tasks;
- code: about 7 hours, covering generation by phase, review and patching to the test environment;
- demo: about 2 hours, covering the scenario, demo data, a run and fixes.
The spec, a detailed technical requirements document, is the part that pays off most. You write the task as free text, the author agent assembles a document, the reviewer checks it, and the loop repeats until the verdict is "OK." Along the way the agent asks questions: what counts as a resolved ticket, whether flow balance can go negative, how to calculate median response time, what target to set for groups with different values. If you agree without reading, the agent will fill gaps in the spec to its own taste, and you will get something other than what you need.
The agent also found discrepancies with the test environment: the description listed three descendants of ITSM Task, but there were actually ten. Waiting statuses and closure codes also differed between groups. Fixing this in the spec is cheaper than reworking a finished widget. The spec, plan, change log and logs are stored in Git, so a disputed calculation can be checked against the original source.
The code was generated by phase: author first, then reviewer, then a patch to the test environment. Review caught small things that are easy to miss by hand, such as an empty default value.
What came out and what it cost
The lead got a new page and a menu item. At the top are three filters: period, support group and task types. Below are four blocks: flow, speed, breakdowns and quality. A separate directory holds the targets. For example, the share of first-line tickets resolved within SLA must not fall below 75%. If the value drops, the tile turns red immediately rather than at the end of the month.
The merge request included 169 records and 161 files, 130 of them localization files. The widget code is about 2 600 lines. An SOP of 288 records went to a clean test environment without manual changes in the interface.
The Opus 5 subscription at $20 a month covered all 786 330 tokens with no limit overruns or extra charges. Corporate scenarios with security requirements more often need an API. There, by the team's calculations, the same volume of responses would have cost $19.66 plus payment for input context, so the difference is small.
Where fixes were needed
Deployment failed three times. The test environment rejected a long Russian tooltip containing an em dash, one field received an empty array instead of a string, and the platform itself had already created a form that the pipeline tried to create again. Each case took about five minutes. The product sync team, which pulls the form into the repository, helped. The demo passed 25 of 28 steps; the remaining three require authorization and manual actions. The run took about 7 minutes and found two defects: an empty block rendered incorrectly, and the tooltips on the target fields were not visible.
The main limitation: nobody read the code line by line, so a human checks the formulas. Some metrics are debatable. The "bottleneck" block is named imprecisely, since SLA violations by priority do not prove a bottleneck. Flow balance goes negative because of tickets waiting for a customer reply, which the team does not control. The report itself is retrospective and does not replace operational dashboards.
What to take away
The team distilled five rules. Answer the agent thoughtfully. Change the spec first, then regenerate the code. Ship by phase and check each one in the test environment. Commit after every patch. Do not skip the demo run: it finds defects that are invisible in the code.
You can start with one page of text: write down the report people request most often, the numbers it needs, the targets and the definition of a "resolved" ticket. Then count how many times it has already been reworked.
- Nov 20, 2026Daydream Launches an AI Stylist That Buys Clothes Itself From 10,000 Brands
- Nov 20, 2026Orca Bots Now Check Your Projects Every 4 Hours on Their Own, With 24 Templates to Start From
- Nov 20, 2026Developer Releases 30 SEO Skills for Claude Code and Cursor, Plus MCP Servers
- Nov 20, 2026Model ML Raises $250 Million for Banker-Focused AI Agents at a $1.5 Billion Valuation
Key Questions Answered
Why the built-in report builder wasn't enough?
SimpleOne's standard reports and dashboards are good at summing up tables. But flow balance (how many tickets came in and how many were closed), median response time, the share of reopened tickets and targets by group all need separate calculation logic. On top of that, the history table, by the team's estimate, grows to roughly 30 million records. As a result, line managers, heads of service departments and IT directors page through ticket lists by hand and miss critical items.
How the four stages work?
SimpleGen uses different agents: a code author, a fixer, a reviewer and others. Time was distributed unevenly across the stages: spec: about 5.5 hours, with 11 "write and review" iterations and 28 questions from the agent; plan: 1.5 hours, with 11 phases and 42 tasks; code: about 7 hours, covering generation by phase, review and patching to the test environment; demo: about 2 hours, covering the scenario, demo data, a run and fixes.
What came out and what it cost?
The lead got a new page and a menu item. At the top are three filters: period, support group and task types. Below are four blocks: flow, speed, breakdowns and quality. A separate directory holds the targets. For example, the share of first-line tickets resolved within SLA must not fall below 75%. If the value drops, the tile turns red immediately rather than at the end of the month. The merge request included 169 records and 161 files, 130 of them localization files. The widget code is about 2 600 lines.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.