Every product on this page was built the way we build for clients: agents execute, objective gates verify, humans judge. The games on Google Play, the agent RPG sandbox, the voice experiments — these are not demos. They are the proving grounds where Reckon’s honesty gates, SkillForge’s skill lifecycle, and Dream Team’s certification were forged on real, shipped code. When you read a metric below, it was measured on these codebases first.
| The Artena Stack — Proven in the Wild |
| Reckon → SkillForge → AgentSpace → Dream Team ↓ ↓ ↓ ↓ Verify → Curate → Coordinate → Certify |
How the Stack Builds Our Own Products
- Agent execution with honest gates. Multi-file edits, migrations, and features are delegated to cost-efficient models through Reckon — and completion is proven by a passing exit gate and the real working-tree diff, never by the model’s “I’m done.” A handed-off task either passes its objective check or comes back marked NOT DONE.
- Skills instead of prompts. The procedures that worked on these codebases — localization in 4,800-line god-files, parity fixes across dual backends, live-acceptance discipline — live as named, gate-proven skills in SkillForge, injected only when their trigger matches, retired when a model outgrows them.
- QA as a standing routine. Dead-code cleanup, abstraction police, orphaned-test scans, and mutation probes run on schedules across our repositories. The fleet maintains the products; engineers review the pull requests.
- Generative assets, human judgment. Where taste matters — icon style, soundtrack mood, level pacing — generated candidates are reviewed by humans. Autonomy is a dial, and on aesthetics we keep it deliberately low.
What Changed When Checks Replaced Promises
- Instructions 0/4. Checks 7/7. In a controlled experiment on our own codebase, asking the executor to “do it right” worked zero times out of four; verifying the result with code worked seven times out of seven. Since then, every requirement in our products is converted, one by one, into a check.
- Two green suites, both lying. A parity campaign found five hidden divergences between dual backends at 1,900 green tests — each suite verified only against itself. The cross-backend parity oracle closed them: final measurement, full agreement 10/10.
- The sealed exam: 19/19. Hidden answers, pre-sealed requirements revealed one by one, verification by execution — the chain passed 19 of 19 for about a dollar, catching 10 of 11 deliberate breaks, six of which were never mentioned in the brief.
- Eleven subsystem migrations, verified live. An external tester checked each migration not by green tests but by independent live behavior checks. Six stable failure patterns became six structural protections; the error class that stalled several tasks in a row simply stopped recurring.
- Seven subsystems and a full-stack messenger, built autonomously. Each time the system studied the reference code, wrote the design, built, tested, and kept the build green — and brought contested calls to human review instead of deciding silently.
Examples of developed applications
Products in the Wild — and the Scanners They Forged
Sky Jumper: Turtles (Google Play) — a retro platformer where a turtle collects stars and defeats aliens. It is also the birthplace of our game-state scanners: a variable accumulated inside a render() loop, a missed reset, a keybind collision, an undo path left untested — these are exactly the temporal bug classes that pass node --check and a render-blind gate, and exactly what Reckon’s structural scanners now catch on every write run. The game stays green; the scanners now protect client code too.
Magic Blocks: Match 3 Games (Google Play) — a match-3 puzzle on floating islands. Its rules engine became our mutation-testing gym: the kind of logic where a weak test passes while a mutant survives, and where the copy-under-test trap (17 green tests, zero coverage) was first caught and made structurally impossible. Level and task generation now runs behind the same honest gates as everything else: generated content ships when the check passes, not when the model says so.
Multiplayer RPGs
AGENT Station 14 — the Agent Sandbox That Became Dream Team
AGENT Station 14, our port of the open-source SS14, began as a multiplayer RPG where NPCs are autonomous agents: they adapt to player behavior, attack, defend, gather resources, hold meaningful conversations, offer quests. It is currently in beta. It is also the environment where our agent-development methodology was born. The SS14 codebase — hundreds of systems, deep simulation logic — is where library-relative trigger weighting learned that words like “station” and “system” are boilerplate in every brief; where localization-first checkpoints were validated on ~4,800-line god-files; where agents first trained alongside real developers. What started as a game sandbox is now the direct precursor of Dream Team: the same environment trains, certifies, and recertifies agents before they touch production code — and production experience flows back into new scenarios.
Apps for Audio Assistant
Voice Agents — an Early Glimpse of AgentSpace
Our voice version of AGENT Station 14 for audio assistants (alpha) lets players run single-player sessions by voice: agents respond to commands, complete quests, and interact with the world in real time. It taught us a lesson that now shapes AgentSpace: in a screenless modality, the agent’s only honest output is a verifiable outcome — so every voice action maps to a game-state change the system can check. Autonomy without a screen forced verification without a screen.
Benefits of using Artena Tech solutions in gaming applications
Why Games Are the Perfect Proving Ground
Games apply the pressure client systems will feel later: real-time state, temporal bugs, reward-hack temptation (an agent under pressure will edit the oracle that judges it), and players who find every edge case. If a verification layer survives a game, it survives a business workflow. Everything on this page now runs on the same stack we sell: Reckon verifies the work, SkillForge curates what the fleet knows, AgentSpace coordinates it, Dream Team certifies it.
Foundations & Related Work
- Agent Skills open standard — agentskills.io: the SKILL.md format behind SkillForge and every learned skill in this portfolio.
- CodeAlive-AI — cross-agent skills, safety hooks, and consilium review patterns used across our stack.
- A formal reward-hacking taxonomy (arXiv:2605.02964) — the basis of our log-replay scanners (answer leakage, forged test markers).
- SpecBench (arXiv:2605.21384) — the hardcoded-answer-table pattern our scanners detect.
- LLM-as-a-Verifier (arXiv:2607.05391) — continuous judge scoring via logit expectation in our no-gate selectors.
- AgentFold (arXiv:2510.24699) — plan-boundary context folding, adopted where it pays and refused where it breaks the prefix-cache discount.
- To Retain or to Adapt (arXiv:2607.05609) — transfer-efficiency memory: proof-of-benefit outranks proof-of-selection.
- STING-style mutation evidence — 77% of SWE-bench-verified instances have a mutant that passes all existing tests; the reason our gates never trust their own tests.