LLMs and natural language processing · AI agents
C3PIO Tracker — developed with AI agents
Our own tracker runs tasks from planning to release with agents, while people set goals and make decisions.
Context and challenge
C3PIO runs several product and research projects in parallel. Planning, coding, review, testing and releases need more capacity. We built a pipeline in which AI agents move tasks from planning to deployment while people set goals and make decisions. These are internal metrics from our own product.
What we built
We designed and built a tracker pipeline. An architect splits work into subtasks, a developer writes code and opens a merge request, a tester checks acceptance criteria, a reviewer reviews code, and auto merge lands the change after green CI and reviewer approval. Stop rules cover budget limits, repeated errors and stage unavailability. Agents use Claude, Codex/GPT and local Qwen. The tracker is developed through this same pipeline.
Results
| Fact | Value | Period / basis | Source |
|---|---|---|---|
| Tracker releases | 31 (0.1.0 → 0.31.0) | 18 Aug–3 Oct 2026 · about 6.5 weeks | DEPLOYED release records, release tags and merge requests in GitLab |
| Tasks in the system | 1,281; 1,080 completed | Since 20 Aug 2026; all tasks in the database | Task records in MERGED, RELEASE, PRODUCTION and DONE states |
| Merged through merge requests | 889 tasks | Since 20 Aug 2026; completed tasks with an MR | Task MR reference and status |
| Automatically merged | 788 tasks | Since 20 Aug 2026; tasks marked auto merged | autoMerged flag: green CI and agent reviewer approval |
| Agent runs | 7,026; 5,745 successful (82%) | Since 20 Aug 2026; all runs | Run records: development 2,926, testing 2,311, review 1,308, planning 400 |
| Speed | Median 1.4 hours from subtask creation to merge | 738 subtasks | Difference between subtask creation and merge timestamps |
| Throughput | 733 tasks merged in 30 days; 82–100 per day | 4 Sep–3 Oct 2026; daily range 29 Sep–3 Oct | Task merge timestamps |
| CI | −62.3% runner minutes; −42.5% pipelines per day | 24h before/after 1 Oct 2026, 03:27 UTC | GitLab API comparison |
| Delivery | Versioned archive and images built automatically for every release | Since version 0.30.0 (3 Oct 2026) | Main pipelines and READY delivery status |
| Scale | 13 active projects, 27 active agents, 41 model profiles | As of 4 Oct 2026 | Project, member and model profile records |
How we counted
Tracker database snapshot from 4 Oct 2026, about 02:00 UTC: releases, tasks, runs, projects, agents and model profiles come from their records. Merge times come from task timestamps; versions and delivery were checked against releases and GitLab. The CI change compares two 24-hour windows around 1 Oct 2026, 03:27 UTC using the GitLab API.
Caveats
- These are internal metrics from our own product, not results for an external customer.
- The pipeline performs approximately 94% of task status transitions automatically. Of 27,771 transitions, 1,643 (5.9%) were made under user accounts, including actions by the owner and an AI operator using the owner’s account. This does not mean no people were involved.
- 82% successful agent runs shows resilience to failures: about one in five runs fails and is retried or stopped by the rules.