AI-native development: agents in production, results measured
We build agentic systems for production with a small team and AI agents, spec- and eval-driven. We measure the result instead of estimating it.
At Staffbite, AI-native development means three things. We build with a small team and AI agents. We deliver agentic systems that do real work in your processes and on your data. And we measure the result: the specification and the evals are the agreement on what counts as done.
AI is already widely used in Finland, where Staffbite is based: 38 % of Finnish enterprises with at least 10 employees used AI in spring 2025, against 20 % across the EU (Figure and source) (Eurostat, 11 December 2025). Getting value from it is harder. In a Gartner survey, only 13 % of CIOs said their enterprise had realised significant or transformational value from AI (Gartner, 19 October 2026).
Request an agent-ready data review
Updated
How we work
Specification and evals first
Before any code, we write a specification: what the system does, what it does not do and under which conditions it is done.
Evals are automated tests that run the system on real example cases. Whenever the model, the prompt or the data changes, they are run again. That way quality is not a matter of opinion.
We measure instead of estimating
A feeling of speed is not a measurement. In METR's controlled trial, experienced open-source developers were 19 % slower (Figure and source) with AI, although afterwards they believed they had been 20 % faster (METR, 10 July 2025).
In newer data, METR estimates that developers were probably sped up, but considers its own evidence for this weak (METR, 24 February 2026). That is why we log durations, tokens and bugs found in our own projects, also when something did not work.
Small batches and a strong foundation
According to the DORA 2025 research, AI amplifies an organisation's existing strengths and weaknesses. The benefits grow with, among other things, small batches, strong version control and high-quality internal platforms (DORA, 23 September 2025). We build on those.
Agents in production
An agentic system does more than answer questions. It retrieves information from systems, draws conclusions and acts within agreed limits: for example, it processes requests for quotation, compiles a report or prepares a decision for a person to approve.
Not every agent project succeeds. Gartner predicts that over 40 % of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear value or inadequate risk controls (Gartner, 25 June 2025). That is why the first version is limited to one process whose value can be measured.
When not to use an agent
- An ordinary program does the task more reliably and more cheaply.
- The data is not available, or its owner is unclear.
- The cost of an error is high, and a person cannot be kept in the decision.
- The agent would at the same time have access to private data, exposure to untrusted content and the ability to send information out. Simon Willison calls this combination the “lethal trifecta” (16 June 2025).
We agree on stop criteria before the pilot: what is measured, where the limit is and when to stop. Only about a third of organisations have such criteria: in a Gartner survey, 31 % had rules for stopping failed pilots (Gartner, 19 October 2026).
Agent-ready data
An agent is only as good as the data it can reach. According to Anthropic, in complex tasks AI is often limited more by access to information than by the model's capability (Anthropic Economic Index, 15 September 2025).
We use this six-point checklist:
- Structured: a machine-readable schema such as JSON or tables, not just PDF files.
- Available: an API or MCP interface, so that the agent can fetch the data itself.
- Documented: the meaning of the fields and the rules of use are written down.
- Permissioned: access control per agent, with only the permissions it needs.
- Up to date: the time of the last update is visible, and there are no stale copies.
- Owned: every data source has a named owner.
The same list works outwards too. When a company's public data is agent-ready, AI assistants find and use its facts. That is what we do in AI visibility.
What you get
- A working agentic system in production, not just a demo.
- Evals that let you see for yourself whether the system works. They stay with you.
- An agent-ready data review: your data is checked against the six-point list, and you get the fixes in order of priority.
- A dependency map: a one-page description of where the data and the AI processing are, whose law applies to them and what the exit plan is. Staffbite's own dependency map is an example.
The architecture is not tied to one language model. Changing the model is a configuration question, not a rewrite, and the evals show whether the new model works as well.
Scopes and durations
We publish durations as ranges. The price is agreed in the offer according to the scope.
- Small: 1–2 weeks, of which 1–2 days of active building. For example, an interactive campaign site, or pre-processing requests for quotation before a person picks them up.
- Medium: 2–4 weeks, of which 4–6 days of active building once the data is ready. For example, a campaign game, a destination map that agents can read, or handling customer service emails.
The examples are imagined requests, not client cases.
Calendar time depends more on the pace of decisions than on the building: how quickly the data, access rights and approvals are in place. map.levifinland.fi and game.levifinland.fi, listed below, each took about four days of active work. One person built them, so there were no approvals to wait for.
The scope depends most on:
- the number of use cases and processes
- the state and availability of the data (see the agent-ready data list above)
- integrations with existing systems
- the number of users and the need for maintenance.
We do not schedule large projects in one go. We build them batch by batch, and the next batch is agreed once the results of the previous one have been measured.
Tell us what you need, and we will make you an offer.
How we built this site
staffbite.fi was built the same way we work for clients: a specification, parallel agents, reviewer agents and automated checks. For every run we log the number of agents, the tokens, the duration and the findings.
For example, implementing the site's visual design on 20 November 2026 took 21 agents, 1.67 million tokens, 38 minutes 44 seconds and two review rounds. Nine valid findings remained after the run, and they were fixed by hand in about 15 minutes. The figures come from Staffbite's build log.
The machine did not see everything. Already at the mock-up stage, the first version of the design rounded the fact highlights over the neighbouring letters. A person noticed it in the browser; the automated check did not. On the other hand, a reviewer agent found a scrollbar bug in the code that only shows for mouse users on Linux and Windows and did not appear in the automated screenshots. That is why a person always looks at the result.
The site's facts are server-rendered text, and machines get the same information in llms.txt, llms-full.txt and sitemap.xml. You can check them with your own agent.
References
- experiencelevi.fi: Chalet Nulkka's accommodation site in Finnish and English. Agents get llms.txt, JSON-LD, an API and MCP (Machine-readable interface).
- karotammela.fi: a personal site and blog in Finnish and English. The same Go and htmx foundation as staffbite.fi, with a Gemini integration.
- map.levifinland.fi: an interactive map of Levi with the real timetables of the ski buses. Self-hosted, and agents can use the data through an API, MCP and GTFS (Machine-readable interface).
- game.levifinland.fi: Skibus Rally, a browser game that uses the map's real route data.
Our own products, Prosessistudio and Kanavoi, are built the same way: Products.
Frequently asked questions
What does AI-native development mean?
A small team and AI agents build agentic systems that do real work in the client's processes. The specification and the evals agree in advance what counts as done, and the result is measured.
Why do AI projects get stuck at the pilot stage?
Often the cause is the data, unclear value or missing rules for stopping. That is why we start from the data and agree on stop criteria before the pilot.
What does eval-driven development mean for AI systems?
Evals are automated tests that run the AI system on real example cases and show whether it works as agreed. They are written before the build and rerun whenever the model, the prompt or the data changes. The evals stay with the client, so the system's quality can be followed after the project too.
Can the language model be changed later?
Yes. The architecture is not tied to one language model. Gemini, Mistral or open weights on your own or a European provider's infrastructure are a configuration question, not a rewrite. The evals show whether the new model works as well.
What does the agent-ready data review include?
We go through your data with a six-point checklist: is it structured, available through an interface, documented, permissioned, up to date and owned. The result is a prioritised list of fixes and an assessment of which process an agent should be brought to first.
What does an AI-native development project cost?
The price depends on the scope, so we do not publish prices. We do publish durations: a small project is done in 1–2 weeks and a medium one in 2–4 weeks. Larger ones are built batch by batch. Tell us what you need, and we will make you an offer.
Get in touch
Start planning now
Tell the AI architect what you are planning. It answers right away and can put together a solution sketch from the conversation that you can share with your team.
- karo@staffbite.fi
Staffbite is based in Helsinki (Location and data). Read about EU-first