Build the world
A hive, flowers, energy, weather, and several virtual bees. The season ends; we count honey, survival, and invalid actions.
OPEN-SOURCE MULTI-AGENT AI LAB
Teach a colony
to survive the unknown.
A deterministic bee simulation where strategies compete on identical worlds, learned policies face predeclared safety gates, and failed experiments stay in the record.
53tests passing
6audit scenarios
Apache 2.0open source
COLONY / 042● LIVE
⬡✦✦✦✦●●●●
HONEY223.68+52.6%
SURVIVAL100%6 / 6
INVALID0.00%safe
NEW HERE? BEGIN BEFORE THE JARGON
A small bee game.
A serious cooperation problem.
Try different minds
Handwritten rules and learned policies solve the same problem: act independently without wasting the colony's shared resources.
Give them the same exam
The same random seed recreates the same world, so a strategy cannot win merely by receiving an easier map.
Publish the evidence
Declared gates, isolated final tests, model hashes, and retained failures make each promoted result traceable.
I AM NEW TO AIUnderstand the game and questionStart with the plain-language guide →
I WORK IN MLInspect training and evidenceOpen the model catalog →
I WORK IN DEVOPS / MLOPSTrace delivery and controlsRead the engineering handoff →
The whole journey: bee simulator → rule baselines → learned policies → fair exams → evidence-backed release.
WHY IT MATTERS BEYOND BEES
A small world for
high-stakes habits.
SweetGold does not sell a bee policy. It makes coordination, fair evaluation, promotion, and audit evidence understandable in one end-to-end system. The workflow may transfer; each industry claim still needs domain validation.
Inspect the complete AI lifecycle
From simulator and baselines to learned policies, isolated final tests, verified models, and auditable decisions.
Shared-resource coordination
Warehouse and robot fleets, drone inspection, mobility research, games, AI evaluation, and technical education.
Teams that teach, test, or govern AI
Universities, corporate academies, ML platforms, governance teams, simulation R&D, and technical consultancies.
ADVANTAGEEvidence is the product
Deterministic comparisons, predeclared gates, retained failures, model integrity, and bilingual explanation form one coherent workflow.
LIMITATIONThe domain is still a proxy
No customer discovery, paid pilot, real fleet integration, multi-user hardening, or measured business outcome exists yet.
01Interview
Validate one repeated, urgent buyer problem.
02Transfer
Connect one external or customer simulator.
03Pilot
Measure time, defect, or audit reduction.
04Productize
Standardize only what repeats.
COMPETEOwn evaluation and promotion
Integrate PettingZoo, RLlib, Isaac Lab, and model registries. Do not compete on environment count, algorithm breadth, GPU scale, or generic tracking.
OPEN-SOURCE REALITYThe code may be copied
Apache-2.0 permits compliant use, modification, redistribution, and commercial reuse. The defense is trusted upstream, brand, adapters, evidence, community, and customer context.
NEXT FEATUREProve transfer first
After customer interviews, connect one PettingZoo environment and publish a stable evidence schema before adding more algorithms or a SaaS dashboard.
Most credible first paths: education and enablement, then an adapter-based evaluation and promotion toolkit. Direct industrial control remains a hypothesis until domain-specific evidence exists.
Read the full customer, SWOT, commercialization, and roadmap analysis →PRODUCT WORKFLOW · M15
Evidence in.
Auditable decision out.
balanced · yield · safety
survival · invalid actions
JSON · Markdown
recommend or reject explicitly
M15 selects among existing strategies; it does not train a new policy. M16 is the latest engineering milestone.
QUESTIONS BEFORE CLAIMS
Start with the
important distinctions.
Is SweetGold just a bee game?
No. The bee world is a reference environment for multi-agent coordination. The durable product is the reproducible evaluation and evidence workflow.
Does one Arena win prove robustness?
No. A small league is a demonstration. Formal claims require predeclared gates, isolated final seeds, and cross-scenario evidence.
Can it control real robots today?
No. The workflow may transfer, but every physical domain still needs its own simulator, integration, safety, and regulatory validation.
Answers for visitors, learners, researchers, engineers, contributors, partners, and governance teams.
Explore the FAQ →LATEST PROMOTED POLICY · M14
Formal performance
of the newest policy.
148.47%median honey vs. Assignment
101.16%worst-case yieldscarce nectar
100%minimum survivalevery scenario
0%invalid actionsevery scenario
M14 · 50 untouched seeds in each of 6 scenarios · selected only on disjoint validation seeds · M15 does not change these results
STRATEGY ARENA
Same world.
Different minds.
Run two policies from the same seed. Compare honey, survival, efficiency and invalid actions live—then replay every frame.
- 01 Matched-seed tournaments
- 02 Verified model registry
- 03 Auditable agent recommendations
MATCH 04 / SEED 202607SEASON 86%
AHierarchical CTDE
218BAssignment
146Δ HONEY+72ALL BEES ALIVE
ENGINEERING GUARANTEES
Research rigor,
built into the system.
Deterministic execution
Seeded environments and matched episodes make every comparison repeatable.
Isolated evaluation data
Preflight checks block leakage across training, validation, selection and final tests.
Policy-as-code gates
Confidence, yield and safety thresholds live in versioned experiment configs.
Verified model supply chain
Immutable releases, exact sizes, SHA-256, licenses and atomic installs protect checkpoints.
Traceable artifacts
JSON run bundles retain commits, runtimes, seed manifests, metrics and decision links.
Automated regression defense
Core tests and end-to-end ML smoke pipelines run in CI; failed audits stay visible.
Enterprise-style engineering practices with an explicit boundary: the built-in server is a local experimentation interface, not a hardened multi-user production service.
LATEST ENGINEERING MILESTONE · M16CPU, Apple MPS, and NVIDIA CUDA are explicit execution targets.
On the measured small behavior-cloning workload, the M1 Pro CPU was about 6.4× faster than MPS because accelerator overhead outweighed useful parallel work. Hardware evidence—not “GPU is always faster”—drives the decision.
Read the hardware benchmark →15 MILESTONES, ONE AUDIT TRAIL
The failures built the policy.
Imitation
Behavior cloning + DAgger
Decentralize
Local actors, global critic
Fail honestly
Robustness gaps preserved
Pass the gates
Hierarchical return control
Decide with evidence
Auditable agent workflow
M10–M12 failed their declared gates. SweetGold records why instead of moving the goalposts.
M14 is the latest promoted policy. M15 is the latest completed workflow: it turns Arena evidence into a constrained, auditable recommendation.
ZERO-DEPENDENCY START
From clone to colony
in three commands.
sweetgold — zsh
$ git clone https://github.com/alanthssss/sweetgold.git
$ cd sweetgold
$ python3 main.py play --port 8080
✓ Arena ready at http://127.0.0.1:8080