Lead-first
The team forms after understanding the project rather than from a rigid template.
OPEN LAB · FLAGSHIP
AI Teams turns multiple agents into an accountable system.
My open-source flagship investigates one practical question: when a software-agent team improves the outcome, and when it only adds coordination, cost and new ways to fail.
Lab / Reviewable delivery
Design · development · verification
Build. Verify. Deliver.
01 · DESIGN
Define the inputBRIEF / Sample document
Expected result defined
Expected result defined
Define exceptionsRULES / Missing fields
02 · BUILD
Extraction and validationDEVELOPMENT / Extraction
Reviewable local delivery
Reviewable local delivery
Prepare the reportDOCUMENTATION / Sources
03 · VERIFICATION
Compare with the originalQA / Amounts and references
Human approvalPENDING / Hold publication
Sample data
AI TEAMS · ACTIVE DEVELOPMENT · WINDOWS VERIFIED
A control plane, not a collection of personas.
A Lead understands the objective, forms the team justified by risk and preserves issues, runs, costs, reviews and blockers in SQLite. Hard gates and evidence decide whether work is accepted, reopened or blocked. The repository is public, installation is verified on Windows and no stable release exists yet.
02
What makes AI Teams different
Dynamic hiring
Engineer, Reviewer, QA and other roles appear only when work and risk justify them.
Durable state
Issues, runs, wakeups, interactions, costs and evidence survive in SQLite and can recover.
Adapters and budgets
Runtime, authentication, health, authority and cost remain explicit for every agent.
Comparative evaluation
A direct agent is compared with solo_lead, lead_quorum and full_team on the same case.
Visible limits
Inconclusive trials and cases where additional agents do not help remain part of the evidence.
03
From objective to verifiable closure
Understand
The Lead translates the project into issues, dependencies, criteria and an exit condition.
→Form
It hires the smallest structure capable of owning the work and its reviews.
→Execute
Heartbeats, runs and budgets keep activity and cost observable.
→Verify
Deterministic tests, review and evidence allow work to be accepted, reopened or blocked.
PUBLIC CODE + EVIDENCE
The flagship publishes its boundaries too
AI Teams 0.1.0 is public under Apache-2.0 with a reproducible Windows clean-room acceptance. Linux and macOS remain unverified; it does not promise full autonomy, guaranteed savings or full-team superiority. RogueBall remains a separate audited Godot runtime case.
A demonstration becomes valuable when it is applied or transferred.
Consulting & development
Turn a prioritized opportunity into an owned, measurable and transferable system.
Explore this route↗Applied training
Give the team the judgment and autonomy to operate, evaluate and extend what was built.
Explore this route↗04
FAQ
Can I use AI Teams now?+
You can inspect and experiment with the repository, but it remains active development without a stable release or verified multi-platform support.
Do more agents produce better software?+
Not necessarily. That is one of the hypotheses the benchmark compares against a direct agent.
What is recorded?+
Work, dependencies, runs, wakeups, interactions, costs, reviews and closure evidence in durable state.
Does RogueBall disappear?+
No. It remains published as an audited Godot build, runtime and verification case with explicit provenance limits.
NEXT STEP
Inspect the system before deciding you need a team.
The flagship page exposes architecture, evaluation, code and limits. If your problem needs similar orchestration, we can start from its closure criteria.
INFORMATION ONLY · NO DATA SUBMISSION
This request is not open during launch. The information remains public to help you prepare the project; if you need consulting or business training, you can request a fit call.