Challenge
An SII team develops and runs a long-lived business application for a car manufacturer - a classic brownfield project. Tickets, from bug fixes to new features, were implemented by hand and estimated in story points. The question for our field test: how much effort does a coding agent save in a brownfield project like this, without compromising quality or accountability?
Approach
Not an autonomous agent but guided agentic development: the engineer works in dialogue with the agent and steers every step.
- Agent environment for ticket work - system prompt, skills for bug fixing, logging and CI, connections to ticket system and code repository.
- From ticket to pull request in dialogue - the agent reads ticket and logs, explores the code base, proposes a plan and implements it; the engineer discusses, corrects and asks for rework.
- Checks before review - every change passes unit tests or an error-free build before a pull request is created.
- Humans decide - engineers review every pull request and decide on the merge.
- Measurement - for each ticket, effort without the agent (story-point estimate) and with the agent (engineers’ own estimate) was recorded.
How the collaboration worked
Result
- 14 tickets: 162 hours estimated manually → 64 hours with the agent, as estimated by the engineers (≈ -60%).
- Small, standardised tickets: ≈ -93%. Large, complex tickets: ≈ -18%. The effect depends heavily on the ticket.
- No change without passing tests and human review.
Results per ticket
| Type | SP | manual (h) | with agent (h) | saving |
|---|---|---|---|---|
| Bug | 2 | 5.6 | 0.25 | 96% |
| Bug | 3 | 8.4 | 0.5 | 94% |
| Bug | 2 | 5.6 | 0.25 | 96% |
| Bug | 3 | 8.4 | 0.5 | 94% |
| Feature | 8 | 22.4 | 10 | 55% |
| Task | 3 | 8.4 | 1 | 88% |
| Feature | 5 | 14 | 1 | 93% |
| Bug | 3 | 8.4 | 0.75 | 91% |
| Task | 2 | 5.6 | 0.5 | 91% |
| Bug | 2 | 5.6 | 0.5 | 91% |
| Improvement | 2 | 5.6 | 0.25 | 96% |
| Improvement | 2 | 5.6 | 0.5 | 91% |
| Feature | 8 | 22.4 | 18 | 20% |
| Improvement | 13 | 36.4 | 30 | 18% |
| Total | 58 | 162.4 | 64 | 60.6% |
Field test in one sprint (March 2026) in one of our own projects, 14 tickets of mixed type and size. Guided work in dialogue, not an autonomous agent. Manual effort from story-point estimates; effort with the agent self-assessed by the engineers. A small sample with estimated shares - a practice experiment, not a study. Models have improved since.
* Field test: measured in one of our own projects
