How we built staffbite.fi with AI agents
staffbite.fi was built with AI agents, and every run was logged: agents, tokens, durations and findings. Including what the machine did not notice.
We rebuilt staffbite.fi with Go and htmx, and AI agents did many of the work stages. For every run we logged the number of agents, the tokens, the duration, the review rounds and the findings. Here are the figures and what we learned from them.
In short: 126 agent runs and about 11.6 million tokens on 20–21 November 2026. A human noticed things the machine did not, and the other way round.
How the work was done
The agents ran in Claude Code with our own plugin, which divides the work in three ways:
- Council: several agents make proposals from different angles, and judge agents score them blind. We used this for decisions that have no automated test, such as the navigation and the URL model.
- Build, review, fix: agents build in parallel in their own lanes, reviewer agents check the work, and the lanes fix it until the reviewers approve.
- Best of n: the same task is done several times in separate git worktrees, and an automated check and a judge pick the best one.
Each builder has its own lane, for example the user interface, the server, the documentation or the content, and it may not edit other lanes' files. The boundaries held even when several runs were going on at the same time.
The figures
| Work | Method | Agents | Tokens | Duration |
|---|---|---|---|---|
| Navigation, site map and URL model | council | 7 | 574,537 | 10 min 6 s |
| Batch 1: URL model, routes, header and footer | build, review, fix | 23 + 2 reviewers | 1,197,337 | 19 min 47 s |
| Content package and publishing rules | best of three | 4 | 446,729 | 10 min 0 s |
| Batch 2: sitemap, robots.txt, llms.txt, JSON-LD | build, review, fix | 11 | 733,106 | 14 min 48 s |
| Design mock-ups A, B and C | 3 parallel agents | 3 | about 505,000 | 11–13 min |
| Study of 113 award-winning sites, mock-ups D and E | 1 + 2 agents | 3 | 706,768 | 28 min + 19–23 min |
| Mock-ups F and G | 2 agents, 2 feedback rounds | 2 | 672,817 | F 47 min, G 22 min |
| Batch 3: design F on the site | build, review, fix | 21 | 1,670,327 | 38 min 44 s |
| Fact bases for the page texts | 2 research agents | 2 | 406,584 | about 10 min |
| Batch 4: texts for eight pages, fact highlights, llms-full.txt | build, review, fix | 25 | 2,757,055 | 35 min 55 s |
| The News section you are reading this in | build, review, fix | 23 | 1,881,915 | 30 min 51 s |
The largest single run was connecting the page texts to the content package (batch 4): 25 agents, four lanes, seven reviewers, two review rounds and 2.76 million tokens. In the first round the reviewers found 6 blocking and 24 minor issues, in the second 2 blocking and 25 minor ones. In the design implementation (batch 3), the nine findings that remained were all valid, and they were fixed by hand in about 15 minutes.
Caught by a human, not the machine
Highlights over the letters. The first mock-up of design F (Faktaruudukko, “fact grid”) rounded the fact highlights outwards to whole grid cells, so they covered neighbouring letters. The site's owner noticed it in the browser; the automated check did not. The feedback was passed to the running agent, which kept its context, fixed the bug and finally ran its own checks for overlap and grid alignment.
An outdated word. The reviewers of batch 2 did not notice that the contact page's description still said “appointment booking”, although there is no appointment booking. It was found when the server was started and the pages were gone through.
Two flaws in the winner. Of the three implementations of the content package, the best one had a gap in the image path check and an incorrect code comment. The judge agent mentioned both, and they were fixed by hand.
Caught by the machine, not a human
An old bug. The council that worked on the navigation found a bug in the code of the time: the hreflang links of the language versions ended up on a 404 page. The finding was verified in the code.
A scrollbar the screenshot does not show. A reviewer in batch 3 found a bug in the code that only shows for mouse users on Linux and Windows. It did not appear in the automated screenshots.
An empty page in the navigation. A reviewer in batch 4 noticed that publishing the first post would have added a News link to the navigation, leading to a page that did not yet have a list. News was kept hidden until the list and the post pages were ready.
A reference without permission. The content reviewer pointed out that the text of a reference had been written before the site's owner had been asked for permission. Permission was asked before publishing.
A keyboard user dropped to the top of the page. In the first version of the News list, pagination changed the list without a page load, but the keyboard focus went back to the top of the page. The UX reviewer marked it as blocking, and it was fixed before the work was approved.
Common misconceptions. The agents compiling the fact bases found claims that are widely repeated but no longer hold. For example, the fine imposed on OpenAI by the Italian data protection authority was annulled on 18 March 2026, and the AI Act's obligations for high-risk systems moved to 2 December 2027.
What did not work
- In the batch 1 run, 8 of the 23 agents ended in an error, because the project's own reviewers were created in the middle of the session and the workflow did not find them. They were run separately.
- Three parallel mock-up agents shared the same working folder and overwrote each other's helper scripts. This had no effect on the repository, and in later runs each agent got its own folder.
- Batch 3 hit its limit of two rounds, and the run's checklist failed because the plan was edited during the run. The work itself was fine.
- Some sources, such as Gartner and UpCloud, blocked automated retrieval. Their figures are marked as less verified in the fact base, and they are checked before use.
What we do not know yet
We have not measured how long the same work would have taken without agents, so we do not claim a figure for speed. A feeling is not enough: in METR's controlled trial, experienced developers were 19 % slower with AI, although they believed they had been 20 % faster (METR, 10 July 2025).
We will keep logging the runs, and comparison runs without the plugin will be marked separately.
What this means for clients
The same approach is used in client projects: a specification, parallel agents, reviewer agents, a human check and measured figures. Read more: AI-native development.